Recommendation data rearrangement method and device, computer program product and electronic equipment

Abstract

The invention discloses a rearrangement method and device for recommendation data, a computer program product and electronic equipment. The method comprises the steps that an advertisement recommendation type sorting sequence of advertisement recommendation objects is acquired; according to a preference sequence generation type condition, a sub-position generation type condition and a fine arrangement generation type condition in the heuristic sequence generation conditions, obtaining at least one candidate sequence in a first candidate sequence, a second candidate sequence and a third candidate sequence; performing sequence evaluation on the candidate sequence to obtain an index score corresponding to an estimated index type of a candidate advertisement recommendation object in the candidate sequence; according to the candidate fusion features, obtaining an index score corresponding to the pre-estimated index type; obtaining a sequence score of the candidate sequence according to the index score; according to the method, candidate sequences which meet the rearrangement score requirement and are selected from the sequence score range are determined as the rearrangement sequence of the advertisement recommendation type sorting sequence, so that the problem of data explosion caused by mass sequence combination in the rearrangement process is avoided.

Description

Technical Field

[0001] The present application relates to the fields of computer application technology and artificial intelligence technology, and more particularly to a method and apparatus for rearranging recommendation data. The present application also relates to a method and apparatus for rearranging recommendation data in a catering service application platform, as well as a computer program product, an electronic device, and a computer storage medium. Background Art

[0002] In today's digital world, providing information to users through application software has become a common technology. Therefore, recommending appropriate information to different users and achieving personalized information recommendations has become an important means to improve user experience. Therefore, recommendation systems have become a vital bridge connecting users and businesses.

[0003] Recommendation systems can recommend content services, including organic and advertising services. Sorting recommended content is a key technical step in connecting users with application services. Its core goal is to accurately predict and rank recommendations based on user interests, context, and business objectives within massive amounts of recommendation data.

[0004] Typically, the recommendation system process includes retrieval, coarse ranking, fine ranking, and re-ranking. These two stages exist primarily to optimize the quality of recommendation results at different levels, ensuring that the recommendations presented to users are both relevant to their interests and meet business objectives, thereby achieving an ecological balance across the entire service system. Summary of the Invention

[0005] The present application provides a method for rearranging recommendation data to solve the performance and time-consuming problems caused by sorting recommendation data in the prior art.

[0006] This application provides a method for rearranging recommendation data, comprising: Get the advertisement recommendation type sorting sequence of the advertisement recommendation object; According to the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, obtaining at least one candidate sequence of the advertisement recommendation object, including a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; Performing sequence evaluation on the candidate sequence to obtain an indicator score corresponding to the estimated indicator type of the candidate advertisement recommendation object in the candidate sequence; including: obtaining an estimated click-through rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain a candidate fusion feature; and obtaining an indicator score corresponding to the estimated indicator type based on the candidate fusion feature; Obtaining, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; The candidate sequence that meets the re-arrangement score requirement and is selected from the sequence score range is determined as the re-arrangement sequence of the advertisement recommendation type ranking sequence.

[0007] In some embodiments, the obtaining of the advertisement recommendation objects in the advertisement recommendation type ranking sequence based on the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, and at least one candidate sequence among a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the advertisement recommendation type ranking sequence generation type condition, include: The first candidate sequence is determined by using the estimated click-through rate, estimated click-through conversion rate, payment index, and weight coefficient of the estimated click-through conversion rate of the advertisement recommendation objects in the advertisement recommendation type sorting sequence as preference sequence generation type conditions; Determine the second candidate sequence by using one or more of the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, the context data, the estimated click-through rate, and the estimated conversion rate as position generation type conditions; The advertisement recommendation type ranking sequence is used as the third candidate sequence determined according to the refined ranking type generation condition.

[0008] In some embodiments, determining the second candidate sequence by using one or more of the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, the context data, the estimated click-through rate, and the estimated conversion rate as position-based type generation conditions includes: One or more of the advertisement recommendation type ranking sequence, natural recommendation type ranking sequence, user portrait, context data, estimated click-through rate, and estimated conversion rate are input as input data for position generation type conditions into the position estimation model; The advertising recommendation type sorting sequence, the natural recommendation type sorting sequence, the user portrait, and at least one feature data of the context data in the input data are processed by the embedding layer of the position estimation model to obtain feature vector data; The feature vector data of the advertising recommendation type sorting sequence and the feature vector data of the natural recommendation type sorting sequence are processed by a cross-attention mechanism through the natural sequence encoding module in the position estimation model to obtain the relationship features between the advertising recommendation type sorting sequence and the natural recommendation type sorting sequence; Passing the ranked sequence of advertisement recommendation types through the advertisement sequence encoding module in the position estimation model to extract correlation features within the advertisement sequence; Passing the user portrait and the context data through the portrait / context encoding layer in the position-based estimation model to extract user attribute features and context features; fusing and connecting the estimated click rate, the estimated conversion rate, the relationship feature, the association feature, and one or more of the user attribute feature and the context feature through the position-based estimation model to obtain a fused connection feature; Processing the fused connection features through the hierarchical extraction mechanism of the position-based prediction model to obtain the position-based matrix of the estimated click-through rate and the position-based matrix of the estimated conversion rate; Determine a position matrix of an estimated click-through conversion rate by combining the position matrix of the estimated click-through rate and the position matrix of the estimated conversion rate of the advertisement recommendation object; The second candidate sequence is determined according to the position matrix of the estimated click-through conversion rate.

[0009] In some embodiments, determining the second candidate sequence according to the position matrix of the estimated click-through conversion rate includes: sequentially placing the advertisement recommendation objects in the advertisement recommendation type sorting sequence into positions in the estimated click-through conversion rate position matrix to obtain position scores of the advertisement recommendation objects at the positions; The advertisement recommendation objects are sorted by position scores, and the advertisement recommendation object with the highest score is selected from the sorting results and placed into the position candidate sequence; The position-based candidate sequence is determined as the second candidate sequence.

[0010] In some embodiments, obtaining the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence includes: Obtaining, based on the log data of the advertisement recommendation type sorting sequence, the estimated click rate, the estimated conversion rate, the natural recommendation sequence, and the advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; The step of fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the advertisement recommendation sequence, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features includes: Obtaining internal relationship features of the advertisement recommendation sequences among the advertisement recommendation sequences through a self-attention mechanism; Obtaining internal relationship features of the natural recommendation sequences between the natural recommendation sequences through the self-attention mechanism; Obtaining the correlation characteristics between the natural recommendation sequence and the advertising recommendation sequence through a multi-head target attention mechanism; The internal relationship features of the advertisement recommendation sequence, the internal relationship features of the natural recommendation sequence, the association relationship features, and the features of the candidate sequence are fused to obtain candidate fused features; and obtaining the indicator score corresponding to the estimated indicator type based on the candidate fused features includes: The candidate fusion features are passed through a multi-layer extraction network model to obtain the estimated exposure rate, estimated click rate, and estimated conversion rate of the candidate advertising recommendation objects in the candidate sequence.

[0011] In some embodiments, obtaining, based on the indicator score, a sequence score of the candidate sequence corresponding to the indicator score includes: Determining a click-through conversion rate of the candidate advertisement recommendation object based on the estimated conversion rate and the estimated click-through rate of the candidate advertisement recommendation object in the indicator score; The sequence score of the candidate sequence is determined according to the estimated exposure rate, the estimated click rate, and the payment data and the estimated click-through rate of the candidate advertisement recommendation object.

[0012] In some embodiments, the step of selecting the candidate sequence that meets the re-ranking score requirement from the sequence score range and determining it as the re-ranking sequence of the advertisement recommendation type ranking sequence includes: sorting the candidate sequences according to the sequence scores; The candidate sequence with the largest sequence score in the selected sorting is determined as the rearranged sequence of the advertisement recommendation type sorting sequence.

[0013] The present application also provides a device for rearranging recommended data, comprising: A first acquiring unit is configured to acquire an advertisement recommendation type sorting sequence of advertisement recommendation objects; a second acquisition unit, configured to obtain, based on the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, the advertisement recommendation object in the advertisement recommendation type sorting sequence, and at least one candidate sequence selected from a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; The third acquisition unit is used to perform sequence evaluation on the candidate sequence to obtain the index score corresponding to the estimated index type of the candidate advertising recommendation object in the candidate sequence; it includes: an acquisition subunit, used to obtain the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertising recommendation sequence corresponding to the advertising recommendation type sorting sequence; a fusion subunit, used to fuse the estimated click-through rate, estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and correlation relationship features between the advertising recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; a calculation subunit, used to obtain the index score corresponding to the estimated index type based on the candidate fusion features; a fourth obtaining unit, configured to obtain, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; A determining unit is configured to determine the candidate sequence selected from the sequence score range and meeting the re-arrangement score requirement as a re-arrangement sequence of the advertisement recommendation type sorting sequence.

[0014] This application also provides a method for rearranging recommended data in a catering service application platform, comprising: Obtaining a ranking sequence of advertisement recommendation types based on advertisement recommendation objects provided by the catering service application platform; According to the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, the advertisement recommendation object in the advertisement recommendation type sorting sequence is obtained, and at least one candidate sequence is selected from a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; Performing sequence evaluation on the candidate sequence to obtain an indicator score corresponding to the estimated indicator type of the candidate advertisement recommendation object in the candidate sequence; including: obtaining an estimated click-through rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain a candidate fusion feature; and obtaining an indicator score corresponding to the estimated indicator type based on the candidate fusion feature; Obtaining, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; The candidate sequence that meets the re-arrangement score requirement and is selected from the sequence score range is determined as the re-arrangement sequence of the advertisement recommendation type ranking sequence.

[0015] The present application also provides a computer program product, including: a computer program, which, when executed by a processor, implements the above-mentioned method for rearranging recommended data, or executes the above-mentioned method for rearranging recommended data in the catering service application platform.

[0016] The present application also provides an electronic device, comprising: processor; The memory is used to store a program for processing data generated by an electronic device. When the program is read and executed by the processor, it executes the method for rearranging the recommended data as mentioned above, or executes the method for rearranging the recommended data in the catering service application platform as mentioned above.

[0017] Compared with the prior art, this application has the following advantages: The present application provides a method for rearranging recommendation data. The method can generate the corresponding first candidate sequence, second candidate sequence and third candidate sequence respectively according to the preference sequence generation type condition, position generation type condition and fine sorting generation type condition in the heuristic sequence generation condition in the heuristic sequence generation condition in the heuristic sequence generation stage, thereby avoiding the problem of data explosion caused by the massive sequence combination in the rearrangement process; in the sequence evaluation stage, the estimated click-through rate, the estimated conversion rate, the internal relationship characteristics of the natural recommendation sequence, the characteristics of the candidate sequence, and the correlation relationship characteristics between the advertising recommendation sequence and the natural recommendation sequence can be fused to obtain candidate fusion features and the estimated index type score of the candidate fusion feature. By introducing the exposure probability index score and weighting the exposure probability to the sequence score calculation stage, the sequence evaluation value is made closer to the true value. In addition, by capturing context information in the sequence evaluation stage, a more accurate estimate is obtained, and the context in the subsequent display stage is guaranteed to be consistent with the context in the evaluation stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a method for rearranging recommended data provided by this application.

[0019] Figure 2 This is a schematic diagram of an embodiment of a reordering service call in a method for reordering recommended data provided in this application.

[0020] Figure 3 This is a schematic diagram of the structure of a position estimation model in a method for rearranging recommended data provided in this application.

[0021] Figure 4 This is a schematic diagram of a network structure embodiment of the implementation principle of a method for rearranging recommended data provided by this application.

[0022] Figure 5 This is a structural diagram of a device for rearranging recommended data provided by this application.

[0023] Figure 6 This is a flowchart of a method for rearranging recommended data in a catering service application platform provided in this application.

[0024] Figure 7 This is a structural diagram of a device for rearranging recommended data in a catering service application platform provided in this application.

[0025] Figure 8 This is a structural diagram of an electronic device provided by this application. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the technical solutions of this application, the following clearly and completely describes this application in conjunction with the drawings in the embodiments of this application. However, this application can be implemented in many other ways different from the following description. Therefore, based on the embodiments provided in this application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of this application.

[0027] It should be noted that the terms "first", "second", "third", etc. in the claims, description and drawings of the present application are used to distinguish similar objects and are not used to describe a specific order or sequence. The data used in this way are interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than that illustrated or described in the present application. In addition, the terms "including", "having" and their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] It should be understood that in the embodiments of the present application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the previous and subsequent associated objects are in an "or" relationship. "Including A, B and / or C" means including any one, any two, or any three of A, B, and C.

[0029] It should be understood that in the embodiments of the present application, "B corresponding to A," "B corresponding to A," "A corresponds to B," or "B corresponds to A" means that B is associated with A and B can be determined based on A. Determining B based on A does not mean determining B based solely on A; B can also be determined based on A and / or other information.

[0030] In light of the above background technology, it can be seen that some life-related application software, such as shopping, takeout dining, and audio and video software, can provide users with personalized or adapted content. The accuracy of recommended content must, on the one hand, satisfy the user experience, and on the other hand, maintain a balanced relationship between merchants, platforms, and users. This allows the application software to operate and develop in an environment that takes into account the needs of multiple parties. Therefore, the sorting of recommended content becomes an important technical means before displaying it on electronic devices. In other words, sorting is a core issue of recommendation systems, with the goal of providing users with a well-ordered queue of service information.

[0031] Sorting usually includes rough sorting, fine sorting, and re-ranking. The main purpose of rough sorting is to quickly screen out candidates that are suitable for users, merchants, etc. from a large number of candidate sets to reduce the pressure of subsequent calculations. Fine sorting is a process of further and more accurately sorting the candidates on the basis of rough sorting to achieve the accuracy of recommendations. Re-ranking is a means of integrating more factors, such as merchant factors, user factors, application software platform itself, and other diverse factors, and then adjusting and optimizing the results of fine sorting so that the recommended content displayed by the application platform can be balanced from multiple angles or dimensions. Recommended content usually also includes natural recommendations based on the user's own situation, as well as advertising recommendations based on merchants, platforms, etc.

[0032] However, in the existing technology, during the re-ranking stage, the advertising recommendation content needs to be combined and mixed into the natural recommendation content. This process will generate a large number of combinations, thereby affecting the online processing performance and also causing time-consuming problems.

[0033] In view of this, the present application provides a method for rearranging recommended data, which can avoid the problem of combination data explosion caused by a large amount of combination data generated due to rearrangement, and can meet the rights and interests of users, platforms, and merchants.

[0034] like Figure 1 As shown, Figure 1 This is a flowchart of a method for rearranging recommendation data provided by the present application, the method comprising: Step S101: obtaining a ranking sequence of advertisement recommendation types of advertisement recommendation objects; Step S102: Based on the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, obtaining at least one candidate sequence of the advertisement recommendation object, including a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; Step S103: performing sequence evaluation on the candidate sequence to obtain an indicator score corresponding to the estimated indicator type of the candidate advertisement recommendation object in the candidate sequence; including: obtaining an estimated click-through rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain a candidate fusion feature; and obtaining an indicator score corresponding to the estimated indicator type based on the candidate fusion feature; Step S104: obtaining a sequence score of the candidate sequence corresponding to the indicator score according to the indicator score; Step S105: Determine the candidate sequence that meets the re-ranking score requirement and is selected from the sequence score range as the re-ranking sequence of the advertisement recommendation type ranking sequence.

[0035] The above steps are described in detail below.

[0036] Regarding step S101: obtaining the advertisement recommendation type sorting sequence of the advertisement recommendation object.

[0037] In this embodiment, the rearrangement involved is mainly for the advertising recommendation objects, and the rearrangement is performed based on the result of the above rearrangement based on the refined sorting sequence. Therefore, this step is based on the advertising recommendation type sorting sequence for the advertising recommendation objects as the refined sorting result. This step is the basis for the rearrangement operation, such as Figure 2 As shown, Figure 2 This is a schematic diagram of an embodiment of a reordering service call in a method for reordering recommended data provided in this application.

[0038] Based on the access request to the application service or the search request in the application service, the natural recommendation service and the advertising recommendation service are called respectively to obtain the refined ranking sequence of the natural recommendation objects and the refined ranking sequence of the advertising recommendation objects, and then the mixed arrangement service is called to trigger the re-arrangement service request of the refined ranking sequence of the advertising recommendation objects through the mixed arrangement service.

[0039] Therefore, during the specific implementation of step S101, the advertisement recommendation type sorting sequence of the advertisement recommendation object may be obtained through a re-arrangement service request, and the advertisement recommendation type sorting sequence may be a refined sorting sequence for the advertisement recommendation object. The triggering of the re-arrangement service request may be based on an access request of the application service, or a search request of a search engine in the application service. In combination with the application scenario, it can be understood that the call of the re-arrangement service may be triggered when accessing the application service software, and the call of the re-arrangement service may be triggered when the application service software has been entered and a certain commodity object or merchant is searched in the search box provided in the application service software. It can be seen that the call of the re-arrangement service is applicable to scenarios where there is a need to sort the recommendation data, such as: food application scenarios (including takeout, in-store and other service forms), shopping application scenarios, and of course travel service application scenarios, etc.

[0040] Regarding step S102: according to the preference sequence generation type condition, the position generation type condition and the refined ranking generation type condition in the heuristic sequence generation condition, obtain at least one candidate sequence of the advertising recommendation object, including a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition.

[0041] Heuristic sequence generation refers to the generation of sequences based on heuristic algorithms. A heuristic algorithm is an algorithm that solves complex problems by finding empirical rules or heuristic knowledge. The algorithm searches for the target solution by exploring the solution space of the problem. In this embodiment, three heuristic sequence generation conditions are provided to realize the generation of candidate sequences, thereby providing a basis for finding the target candidate sequence. The three heuristic sequence generation conditions may include: a preference sequence generation type condition, a position generation type condition, and a precision generation type condition. At least one of these three heuristic sequence generation conditions may be adopted in the specific implementation process. The specific implementation process of this step may include: Step S102-1: The predicted click-through rate (pCTR), the predicted click-through conversion rate (pCTCVR), the payment index (BID), and the weight coefficient (ctcvr_weight) of the predicted click-through conversion rate of the advertisement recommendation objects in the advertisement recommendation type sorting sequence are used as preference sequence generation type conditions to determine the first candidate sequence.

[0042] In this embodiment, the determination or generation of the first candidate sequence can be implemented using the following formula: Where a and b represent hyperparameters for preference; pCTR is the estimated click-through rate; bid is the ad bid; pCTCVR is the estimated click-through conversion rate; and w is the weighting factor for the estimated click-through conversion rate. By configuring a and b, you can generate first candidate sequences with different preferences. w can also be configured with different weighting factors.

[0043] Step S102-2: Use one or more of the advertising recommendation type sorting sequence, natural recommendation type sorting sequence, user portrait, context data, estimated click rate, and estimated conversion rate as position generation type conditions to determine the second candidate sequence.

[0044] In this embodiment, the second candidate sequence may be determined or generated in the following manner: Step S102-21: One or more of the ad recommendation type ranking sequence, natural recommendation type ranking sequence, user portrait, context data, estimated click rate, and estimated conversion rate are input as input data for position generation type conditions into the position estimation model; Figure 3 As shown, Figure 3This is a schematic diagram of the structure of a position prediction model in a method for rearranging recommendation data provided in the present application, which includes the estimated click-through rate, estimated conversion rate (pCTR / pCVR, pCTR: Predicted Click-Through Rate, i.e., the probability that a user clicks an ad after seeing it; pCVR: Predicted Conversion Rate, i.e., the probability that a user completes a target behavior after clicking an ad), user profile (Uer Proflie), context data (Context, which can be understood as the context of the refined ranking sequence), advertising recommendation type ranking sequence (Ad Seq can also be called advertising sequence), natural recommendation type ranking sequence (Na Seq can also be called natural sequence). These data can be obtained through user behavior data, advertising log data, etc. One or more of these data are used as input data and input into the position prediction model to obtain the output second candidate sequence. The position prediction model can be constructed based on a list-wise sequential prediction model.

[0045] When the input data is input into the location estimation model, the corresponding processing process includes: Step S102-22: The input data is processed by the embedding layer of the position estimation model using the sorted sequence of the advertising recommendation type, the sorted sequence of the natural recommendation type, the user portrait, and at least one feature data of the context data to obtain feature vector data; Step S102-23: The feature vector data of the ranked sequence of advertising recommendation types and the feature vector data of the ranked sequence of natural recommendation types are processed using a cross-attention mechanism by the natural sequence encoding module in the position estimation model to obtain relationship features between the ranked sequence of advertising recommendation types and the ranked sequence of natural recommendation types. In this embodiment, the ranked sequence of natural recommendation types includes: identifiers of natural recommendation objects, such as product identifiers or merchant identifiers, and related auxiliary information (sideinfo). The features between the natural recommendation sequence and the advertising recommendation sequence are extracted by the natural sequence encoding module (Na Encoder). In this embodiment, the Na Encoder module is primarily based on a network including a multi-head target-attention mechanism, which can also be understood as a cross-attention mechanism. Multi-head target-attention refers to an attention mechanism used in fields such as recommendation systems to dynamically calculate the importance of each item in a behavior sequence based on the target item when processing a user behavior sequence. Cross-Attention is an attention mechanism that establishes a connection between two different sequences. In fact, multi-head target-attention can be seen as an application of cross-attention. The Na Encoder module uses multi-head target-attention to learn the relevant information between the ad and natural recommendation sequences. In other words, it captures the influence of the natural recommendation sequence on the ad recommendation sequence. This allows the Na Encoder module to extract the relationship features between the ad recommendation type ranking sequence and the natural recommendation type ranking sequence.

[0046] Step S102-24: The ad recommendation type ranking sequence is passed through the ad sequence encoding module (Ad Encoder) in the position estimation model to extract relevant features within the ad sequence. The relevant data of the ad recommendation type ranking sequence may include: the ad's item_id (the identifier of the ad recommendation object, such as the product ID or merchant ID) and related auxiliary information (sideinfo). These features can be extracted by the ad sequence encoding module. The Ad Encoder module primarily includes a self-attention mechanism network to learn the relevant information between ad sequences.

[0047] Step S102-25: The user profile and context data are passed through the profile / context encoding layer of the location-based prediction model to extract user attribute features and context features. Context data can be the context of the ad recommendation type ranking sequence, such as time data, query data, and other different types of context data. The user profile can include user attribute data. Features involved in the context data and attribute features involved in the user profile can be extracted through the user and context encoding module.

[0048] Step S102-26: The estimated click-through rate, the estimated conversion rate, the relationship features, the association features, and one or more of the user attribute features and context features are fused and connected through the position-based prediction model to obtain fused connection features. In this embodiment, one of the pCTR / pCVR position-based prediction model input data of the refined ranking sequence in the refined ranking stage is used, and features are extracted separately and directly applied to the Concat & Fusion layer (splicing and fusion layer). This can strengthen the position-based prediction model's focus on the estimated click-through rate and estimated conversion rate (pCTR / pCVR) features, improve the comprehensiveness of the data in the rearrangement stage, and provide a foundation for the accuracy and adaptability of subsequent rearrangement results. Regarding the estimated click-through rate and estimated conversion rate (pCTR / pCVR), in this embodiment, they can be input as a set of data into the position estimation model. Of course, they can also be split into two independent data based on different application scenarios. For example, in the search scenario, the two can be input as a set of data. Because there is a positive correlation between user clicks and conversion behaviors, therefore, when performing position estimation on the sequence, introducing the estimated conversion rate can play a positive guiding role in the estimation of the click-through rate. In addition, by using the estimated click-through rate and the estimated conversion rate (pCTR / pCVR) as a set of data, the consistency of the refined ranking link and the re-ranking link can be guaranteed. Of course, in some other scenarios, the estimated click-through rate and the estimated conversion rate (pCTR / pCVR) can also be input as independent data respectively.

[0049] Step S102-27: Process the fused connection features through the hierarchical extraction mechanism of the position-based prediction model to obtain the position-based matrix of the estimated click-through rate and the position-based matrix of the estimated conversion rate. The hierarchical extraction mechanism can be based on the PLE expert network, and gradually separate the shared features and task-specific features through multi-layer extraction. The PLE expert network can include two layers of CGC networks, and each layer of the CGC network can include 3 specific-experts for CTR tasks and 3 specific-experts for CVR tasks. Output two data corresponding to the CTR task and the CVR task, namely: the position-based matrix of the estimated click-through rate and the position-based matrix of the estimated conversion rate.

[0050] Step S102-28: The estimated click-through rate position matrix and the estimated conversion rate position matrix of the advertising recommendation object are combined to determine the estimated click-through conversion rate position matrix. The estimated click-through rate position matrix and the estimated conversion rate position matrix of the advertising recommendation object are integrated through the fully connected layer, and then the estimated position result of pCTCVR (pCTR×pCVR) is obtained through the activation function (sigmoid). Figure 3 As shown in the figure, slot 1 refers to the estimated value of position 1.

[0051] In this embodiment, the position estimation model can be constructed based on a list-wise network, with list_size representing the length of the candidate sequence and position_num representing the number of positions to be estimated. The dimension of the position matrix can then be expressed as list_size × position_num. When calculating the loss, it is necessary to maintain the position matrix of the estimated click-through conversion rate and the position matrix of the estimated click-through rate. The calculation formula for loss is as follows: Where y represents the sample label (the value of not clicked or not converted to a negative sample is 0, and the value of clicked or converted to a positive sample is 1); i is the i-th ad; k is the k-th position; is the estimated click-through rate or conversion rate of the i-th ad in the k-th position, Indicates the calculation of cross entropy; It is an indicator function. If the condition is met, it takes 1, otherwise it takes 0. That is, it indicates whether the i-th advertisement is exposed at the k-th position. If it is exposed, it takes 1, and if it is not exposed, it takes 0.

[0052] By calculating the loss of each position in the estimated click-through rate position matrix, we can get the loss matrix and the estimated exposure position matrix ( ) calculates the Hadamard product to obtain the loss matrix. This can be understood as masking the unexposed loss using the estimated exposure position matrix. In this embodiment, the loss at the exposure position is used.

[0053] The ESMM multi-task network calculates the cross entropy of CTR (Click-Through Rate) and CTCVR (Click-Through & Conversion Rate) as the final loss, and distributes the weights through λ1 and λ2, as shown in the following formula: in, represents the click-through rate cross entropy, Represents the weight coefficient of click-through rate cross entropy; represents the cross entropy of click-through conversion rate, Represents the weight coefficient of the cross entropy of click-through conversion rate.

[0054] Step S102-29: Determine the second candidate sequence based on the position matrix of the estimated click-through conversion rate. In this embodiment, the specific implementation process of this step may include: Step S102-291: Place the advertisement recommendation objects in the advertisement recommendation type sorting sequence into positions in the estimated click-through conversion rate position matrix in sequence, and obtain the position score of the advertisement recommendation object at the position; in this embodiment, the position score can be calculated using the following formula: in, is the click rate of the rearranged sequence, Bid for your ad, is the click-through conversion rate weight coefficient, Estimated click-through rate.

[0055] Step S102-292: sorting the advertisement recommendation objects by their position scores, and selecting the advertisement recommendation object with the highest score from the results of sorting by position scores and placing it into a candidate sequence for position scores; Step S102-293: Determine the position-based candidate sequence as the second candidate sequence.

[0056] Step S102-3: Using the advertisement recommendation type ranking sequence as the third candidate sequence determined according to the refined ranking type generation condition. In this step, the refined ranking sequence of the obtained advertisement recommendation type ranking sequence can be directly used as the third candidate sequence.

[0057] In this embodiment, three generation type conditions are used as an example to obtain three candidate types. The three candidate sequences corresponding to the three candidate types are used as input data for the sequence evaluation model to perform sequence evaluation, i.e., executing step S103. In other embodiments, one or more of the three generation type conditions may be selected to obtain corresponding candidate sequences. This is not limited to performing all three generation type conditions simultaneously, nor is it limited to the three generation conditions. For example, a rough ranking generation condition may also be included, i.e., a rough ranking sequence for the advertising recommendation object may be directly used as another candidate sequence.

[0058] Regarding step S103: performing sequence evaluation on the candidate sequence to obtain the index scores corresponding to the estimated index types of the candidate advertisement recommendation objects in the candidate sequence.

[0059] The sequence evaluation in step S103 can be processed using a sequence evaluation model. Considering the impact of the relationship between the natural recommendation sequence and the advertising recommendation sequence on re-ranking, the natural recommendation ranking and the advertising recommendation ranking are split into two independent links for corresponding processing. From the perspective of sequence evaluation model modeling, the natural recommendation and advertising recommendation sequences are split into two separate sequences and modeled separately.

[0060] In the process of processing the advertising recommendation sequence, the advertising recommendation sequence is used as one of the input feature data of the sequence evaluation model, and the internal relationship feature data between the advertising recommendation sequences is obtained through the self-attention mechanism.

[0061] In the process of processing natural recommendation sequences, the internal relationship features of natural sequences between natural recommendation sequences are obtained through the self-attention mechanism, and the internal relationship features of natural sequences are used as one of the feature data. The Multi-Head Target Attention mechanism is then used between the advertising recommendation sequence and the natural recommendation sequence to obtain the correlation relationship features between the natural recommendation sequence and the advertising recommendation sequence, and the said correlation relationship features are also used as one of the feature data.

[0062] The pCTR and pCVR features of the ad recommendation sequence are used as input data for the sequence evaluation model and fused with the association and internal relationship features to generate fused features. For example, if the pCTR values ​​in the ad recommendation sequence are [0.3, 0.2, 0.1] and the pCTR values ​​in the natural recommendation sequence are [0.5, 0.05, 0.15, 0.35], then the pCTR values ​​are sorted from smallest to largest [0.05, 0.1, 0.15, 0.2, 0.3, 0.35, 0.5], where the ad recommendation sequence values ​​correspond to the positions [4, 3, 1] (including position 0), and these positions are then passed as features to the sequence evaluation model.

[0063] Of course, the input data of the sequence evaluation model may also include user portraits, context data, etc., and may also include other feature data in different application scenarios. For example, in an instant delivery scenario, it may also include delivery distance, delivery price and other data.

[0064] The PLE layer (expert network layer) generates scores for three metrics: estimated exposure rate, estimated click-through rate, and estimated conversion rate. In this embodiment, these metrics are weighted to a loss value during sequence evaluation model training, thereby facilitating more accurate scores when the sequence evaluation model subsequently evaluates candidate sequences.

[0065] Therefore, the specific implementation process of step S103 may include: The step of obtaining the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence includes: According to the log data of the advertisement recommendation type sorting sequence, the estimated click rate, estimated conversion rate, natural recommendation sequence, and advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence are obtained.

[0066] The step of fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the advertisement recommendation sequence, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features includes: Obtaining internal relationship features of the advertisement recommendation sequences among the advertisement recommendation sequences through a self-attention mechanism; Obtaining internal relationship features of the natural recommendation sequences between the natural recommendation sequences through the self-attention mechanism; Obtaining the correlation characteristics between the natural recommendation sequence and the advertising recommendation sequence through a multi-head target attention mechanism; The internal relationship features of the advertisement recommendation sequence, the internal relationship features of the natural recommendation sequence, the association relationship features, and the features of the candidate sequence are fused to obtain candidate fusion features.

[0067] Obtaining the indicator score corresponding to the estimated indicator type based on the candidate fusion features includes: The candidate fusion features are passed through a multi-layer extraction network model to obtain the estimated exposure rate, estimated click rate, and estimated conversion rate of the candidate advertising recommendation objects in the candidate sequence.

[0068] Of course, this can also include: According to the log data of the advertising recommendation type sorting sequence, user portraits and context data are obtained, and the feature data of the user portraits, the feature data of the context, the internal relationship features of the advertising recommendation sequence, the internal relationship features of the natural recommendation sequence, the association relationship features, and the features of the candidate sequence are fused to obtain candidate fusion features.

[0069] Regarding step S104: according to the index score, the sequence score of the candidate sequence corresponding to the index score is obtained. In this step, the sequence score can be obtained using the following formula: in, represents the estimated exposure rate of the kth advertisement in the i-th candidate sequence; represents the estimated click-through rate of the k-th advertisement in the i-th candidate sequence; represents the bid price of the kth advertisement in the i-th candidate sequence; represents the estimated click-through conversion rate of the kth advertisement in the i-th candidate sequence, and w represents the weight coefficient of the estimated click-through conversion rate.

[0070] Regarding step S105: the candidate sequence that meets the re-arrangement score requirement and is selected from the sequence score range is determined as the re-arrangement sequence of the advertisement recommendation type sorting sequence.

[0071] Step S105 sorts the candidate sequences based on their sequence scores obtained in step S104. The sorting can be either descending or ascending. Regardless of the sorting method, the sorting results and corresponding sequence scores of the candidate sequences are re-ranked according to the highest score among the candidate sequences. That is, the candidate sequence corresponding to the highest score is used as the re-ranked sequence of the advertising recommendation type ranking sequence. Based on the above-mentioned sequence score calculation formula, according to the candidate sequence scores (rankScore), the candidate sequence corresponding to the maximum value (max) is used as the re-ranked sequence.

[0072] The above is a description of the specific execution process of the method for rearranging the recommended data provided by this application. Figure 4 As shown, Figure 4 This is a schematic diagram of a network structure embodiment of the implementation principle of a method for rearranging recommended data provided by this application. It can be seen that the method for rearranging recommended data provided by this application can be based on Figure 4 The network structure shown in this embodiment can be implemented by including a heuristic sequence generation model and a sequence evaluation model. The heuristic sequence generation model includes a sequence generation module and a sequence selection module. The sequence evaluation module includes an embedding layer, an encoding layer, a fusion layer, and an expert layer. Based on the network structure in this embodiment and the above content, the implementation principles of the recommended data rearrangement method in this application are briefly described.

[0073] The advertisement recommendation type ranking sequence, that is, the advertisement refined ranking sequence, is input as input data into the heuristic sequence generation model. Through the sequence generation module in the heuristic sequence generation model, the corresponding first candidate sequence, second candidate sequence, and third candidate sequence are generated respectively according to the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition. The data involved in the preference sequence generation type condition include: w weight, that is, ctcvr_weight coefficient weight, Bid index, ctr index, etc. The values ​​of these indexes can be set according to the actual application scenario and the specific requirements of the sorting. Specific numerical values ​​are not given as examples here, because different application scenarios and sorting requirements can correspond to different index values. The corresponding candidate sequence can be obtained according to the position generation model in the position generation type condition, which can be implemented by a greedy algorithm. The refined ranking generation type condition can directly use the refined ranking sequence of advertisement recommendation as the candidate sequence.

[0074] After generating candidate sequences based on the three generation conditions above, the candidate sequences, the pCTR and pCVR of the ad recommendation type ranking sequence (i.e., the existing ad refined ranking sequence), the ad recommendation type ranking sequence, and the organic recommendation type ranking sequence (i.e., the existing organic refined ranking sequence) are processed as input data for the sequence evaluation model. The goal is to obtain values ​​for the candidate sequences along three dimensions: pexp (estimated exposure rate), pctr (estimated click-through rate), and pcvr (estimated conversion rate). Specifically, the pexp, pctr, and pcvr values ​​for the first candidate sequence, the pexp, pctr, and pcvr values ​​for the second candidate sequence, and the pexp, pctr, and pcvr values ​​for the third candidate sequence. Therefore, the sequence evaluation model includes modeling based on organic-advertising sequences and multi-objective modeling based on exposure probability. This processing not only considers the internal feature relationships between the organic recommendation type ranking sequence and the ad recommendation type ranking sequence itself, but also the relationship features between the ad recommendation type ranking sequence and the ad recommendation type ranking sequence. By introducing exposure probability and weighting pexp, pctr, and pcvr, the subsequent sequence score calculation becomes more reasonable. This is because the exposure of recommended data affects user behavior. Therefore, introducing exposure probability and weights can avoid deviations in sequence score calculation.

[0075] Finally, the pexp, pctr, and pcvr corresponding to the candidate sequences are calculated through the sequence selection module in the heuristic generation sequence to obtain the scores corresponding to the candidate sequences. The candidate sequences are sorted according to the scores, and the candidate sequences with the highest scores in the sorting are selected as the rearranged sequences of the advertising recommendation type sorting sequence and output.

[0076] In this embodiment, the position-based prediction model and the sequence evaluation model have a similar network structure, the difference being the difference in input data and output data. The position-based prediction model outputs position information, and the second candidate sequence is a candidate sequence obtained based on the position information. For example, the position-based pCTR and pCVR of the advertising recommendation object are obtained, all advertising recommendation objects are placed in the pre-selection queue, the first position is selected to traverse and calculate the rankscore of all advertising recommendations (the formula in step S102), and they are sorted according to the rankscore, and those with scores that meet the requirements are selected and placed in the candidate result queue, and then removed from the pre-selection queue. Repeat the above operation for the next position until the pre-selection queue is empty, and the candidate queue in the candidate result queue is the second candidate sequence, i.e., the position-based candidate sequence. Please refer to the above step S102 for details, which will not be repeated here.

[0077] When modeling multi-targets based on exposure probability, the training process can select exposed samples as positive samples and unexposed samples as negative samples. Using multi-task modeling combined with ESMM, the ctr, cvr, and exp probabilities are estimated. The loss function is weighted, as shown in the following formula: in, represents the estimated exposure loss ( ), Represents the loss of exposure click rate ( ), Represents the loss of exposure-click conversion rate ( ) weight coefficient; pexp represents the estimated exposure rate; expctr represents the estimated exposure click-through rate, which is the product of the estimated exposure rate and the estimated click-through rate; expctcvr represents the estimated exposure click-through rate, which is the product of the estimated exposure rate, the estimated click-through rate, and the estimated conversion rate.

[0078] The above is a summary description of a method for rearranging recommended data provided by the present application in combination with the network model implementation principle. From the above, it can be seen that the method for rearranging recommended data provided by the present application can generate the required candidate sequences according to the heuristic conditions in at least one embodiment provided by the heuristic sequence generation stage, thereby avoiding the problem of data explosion caused by the massive sequence combinations caused in the rearrangement process; the sequence evaluation stage can introduce the exposure probability and weight the exposure probability to the sequence selection score calculation stage, so that the sequence evaluation value is closer to the true value. In addition, by capturing contextual information in the sequence evaluation stage, a more accurate estimate can be obtained, and the context in the subsequent display stage can be guaranteed to be consistent with the context in the evaluation stage.

[0079] The above is a detailed description of an embodiment of a method for rearranging recommended data provided by this application. Corresponding to the embodiment of a method for rearranging recommended data provided above, this application also discloses an embodiment of a device for rearranging recommended data. Figure 5 Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For relevant details, please refer to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0080] like Figure 5 As shown, Figure 5 : is a schematic diagram of the structure of a device for rearranging recommended data provided by this application, the device comprising: The first acquisition unit 501 is used to acquire an advertisement recommendation type ranking sequence of an advertisement recommendation object; The second acquisition unit 502 is configured to obtain, based on the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, at least one candidate sequence of the advertisement recommendation object, including a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; The third acquisition unit 503 is used to perform sequence evaluation on the candidate sequence to obtain the index score corresponding to the estimated index type of the candidate advertising recommendation object in the candidate sequence; the third acquisition unit includes: an acquisition subunit, used to obtain the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertising recommendation sequence corresponding to the advertising recommendation type sorting sequence; a fusion subunit, used to fuse the estimated click-through rate, estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and correlation relationship features between the advertising recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; a calculation subunit, used to obtain the index score corresponding to the estimated index type based on the candidate fusion features; A fourth obtaining unit 504 is configured to obtain, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; The determining unit 505 is configured to determine the candidate sequence selected from the sequence score range and meeting the re-arrangement score requirement as the re-arrangement sequence of the advertisement recommendation type sorting sequence.

[0081] For the specific content of the first acquisition unit 501, please refer to the above step S101 and will not be described in detail here.

[0082] The specific implementation process of the second acquiring unit 502 may include: a first determining subunit, a second determining subunit, and a third determining subunit.

[0083] The first determination subunit is used to use the estimated click-through rate, estimated click-through conversion rate, payment index, and weight coefficient of the estimated click-through conversion rate of the advertising recommendation object in the advertising recommendation type sorting sequence as preference sequence generation type conditions to determine the first candidate sequence.

[0084] The second determination sub-unit is used to use one or more of the advertising recommendation type sorting sequence, natural recommendation type sorting sequence, user portrait, context data, estimated click-through rate, and estimated conversion rate as position generation type conditions to determine the second candidate sequence.

[0085] The third determining subunit is configured to use the advertisement recommendation type ranking sequence as the third candidate sequence determined according to the refined ranking type generation condition.

[0086] The specific implementation process of the second determining subunit may include: An input subunit, configured to input one or more of the advertising recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, the context data, the estimated click-through rate, and the estimated conversion rate as input data for the position generation type condition into the position estimation model; a processing subunit configured to process at least one feature data of the advertising recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, and the context data in the input data through the embedding layer of the position estimation model to obtain feature vector data; A first feature acquisition subunit is configured to perform cross-attention processing on the feature vector data of the advertising recommendation type sorting sequence and the feature vector data of the natural recommendation type sorting sequence through the natural sequence encoding module in the position estimation model to obtain relationship features between the advertising recommendation type sorting sequence and the natural recommendation type sorting sequence; A second feature acquisition subunit is configured to pass the advertisement recommendation type ranking sequence through the advertisement sequence encoding module in the position estimation model to extract associated features within the advertisement sequence; A third feature acquisition subunit is configured to extract user attribute features and context features by passing the user portrait and the context data through a portrait / context encoding layer in the position estimation model; a fusion subunit, configured to fuse and connect the estimated click-through rate, the estimated conversion rate, the relationship feature, the association feature, and one or more of the user attribute feature and the context feature through the position-based estimation model to obtain a fused connection feature; A first sub-position determination sub-unit is configured to process the fused connection features through a hierarchical extraction mechanism of the sub-position estimation model to obtain a sub-position matrix of the estimated click-through rate and a sub-position matrix of the estimated conversion rate; A second position determination subunit is configured to determine a position matrix of an estimated click-through conversion rate by merging the position matrix of the estimated click-through rate and the position matrix of the estimated conversion rate of the advertisement recommendation object; The second candidate sequence determination subunit is configured to determine the second candidate sequence according to the position matrix of the estimated click-through conversion rate.

[0087] The specific implementation process of the second candidate sequence determination subunit may include: a position score acquisition subunit, configured to sequentially place the advertisement recommendation objects in the advertisement recommendation type sorting sequence into positions in the position matrix of the estimated click-through conversion rate, and obtain position scores of the advertisement recommendation objects at the positions; a selection subunit, configured to select the advertisement recommendation object with the highest score from the ranking result of ranking the advertisement recommendation objects according to their position scores and place it into a candidate sequence for the position; A sequence determination subunit is configured to determine the position-based candidate sequence as the second candidate sequence.

[0088] The acquisition subunit in the third acquisition unit 503 is specifically used to obtain the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertising recommendation sequence corresponding to the advertising recommendation type sorting sequence based on the log data of the advertising recommendation type sorting sequence. The fusion subunit is specifically used to obtain the internal relationship features of the advertising recommendation sequences between the advertising recommendation sequences through the self-attention mechanism; obtain the internal relationship features of the natural recommendation sequences between the natural recommendation sequences through the self-attention mechanism; obtain the correlation relationship features between the natural recommendation sequences and the advertising recommendation sequences through the multi-head target attention mechanism; fuse the internal relationship features of the advertising recommendation sequences, the internal relationship features of the natural recommendation sequences, the correlation relationship features, and the features of the candidate sequences to obtain candidate fusion features. The calculation subunit is specifically used to obtain the estimated exposure rate, estimated click-through rate, and estimated conversion rate of the candidate advertising recommendation objects in the candidate sequence through a multi-layer extraction network model.

[0089] The specific implementation process of the fourth acquisition unit 504 may include: a first determination subunit, used to determine the click-through conversion rate of the candidate advertising recommendation object based on the estimated conversion rate and estimated click-through rate of the candidate advertising recommendation object in the indicator score; a second determination subunit, used to determine the sequence score of the candidate sequence based on the estimated exposure rate, the estimated click-through rate, and the payment data and estimated click-through rate of the candidate advertising recommendation object.

[0090] The specific implementation process of the determination unit 505 includes: a sorting subunit, which is used to sort the candidate sequences according to the sequence scores; and a determination subunit, which is used to determine the candidate sequence with the largest sequence score in the sorting as the rearranged sequence of the advertisement recommendation type sorting sequence.

[0091] The above is a brief description of an embodiment of a device for rearranging recommended data provided in this application. For the specific implementation process of the device, please refer to the relevant content of the above method embodiment.

[0092] Based on the above, this application also provides a method for rearranging recommended data in a catering service application platform, such as Figure 6 As shown, Figure 6 This is a flowchart of a method for rearranging recommended data in a catering service application platform provided by this application. The method is mainly described using a catering service application platform or application software as an example, that is, the implementation method of rearranging recommended data is explained using catering services as an application scenario. Catering services can be instant services such as takeout and retail. Of course, it is not limited to this application scenario. The method includes: Step S601: obtaining a ranking sequence of advertisement recommendation types based on advertisement recommendation objects provided by a catering service application platform; Step S602: obtaining the advertisement recommendation objects in the advertisement recommendation type sorting sequence according to the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, and at least one candidate sequence selected from a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; Step S603: performing sequence evaluation on the candidate sequence to obtain the index score corresponding to the estimated index type of the candidate advertisement recommendation object in the candidate sequence; including: obtaining the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; fusing the estimated click-through rate, estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and obtaining the index score corresponding to the estimated index type based on the candidate fusion features; Step S604: obtaining a sequence score of the candidate sequence corresponding to the indicator score according to the indicator score; Step S605: Determine the candidate sequence that meets the re-ranking score requirement and is selected from the sequence score range as the re-ranking sequence of the advertisement recommendation type sorting sequence.

[0093] For the specific contents of the above steps S601 to S605, please refer to the contents of the above steps S101 to S105, which will not be described in detail here.

[0094] Accordingly, the present application also provides a device for rearranging recommended data in a catering service application platform, such as Figure 7 As shown, Figure 7 This is a schematic diagram of the structure of a device for rearranging recommended data in a catering service application platform provided by this application; the device includes: The first acquisition unit 701 is configured to acquire an advertisement recommendation type ranking sequence based on advertisement recommendation objects provided by the catering service application platform; A second acquisition unit 702 is configured to obtain, based on the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, the advertisement recommendation object in the advertisement recommendation type sorting sequence, and at least one candidate sequence selected from a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition. The third acquisition unit 703 is used to perform sequence evaluation on the candidate sequence to obtain the index score corresponding to the estimated index type of the candidate advertisement recommendation object in the candidate sequence; it includes: an acquisition subunit, used to obtain the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; a fusion subunit, used to fuse the estimated click-through rate, estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; a calculation subunit, used to obtain the index score corresponding to the estimated index type based on the candidate fusion features; A fourth obtaining unit 704 is configured to obtain, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; The determining unit 705 is configured to determine the candidate sequence selected from the sequence score range and meeting the re-arrangement score requirement as a re-arrangement sequence of the advertisement recommendation type sorting sequence.

[0095] The above technical solutions of the present application can be applied to the transaction and delivery services of instant e-commerce platforms, such as Taobao Flash Purchase, Taoxianda, Ele.me takeout and retail, etc. Based on the above content, the present application also provides a computer program product, including: a computer program, which, when executed by a processor, implements the relevant content of the above-mentioned method for rearranging recommended data, or executes the relevant content of the method for rearranging recommended data in the above-mentioned catering service application platform.

[0096] Based on the above content, the present application also provides an electronic device, such as Figure 8 As shown, Figure 8 The present application provides a schematic structural diagram of an electronic device, which includes: Processor 801; Memory 802 is used to store a program for processing data generated by an electronic device. When the program is read and executed by the processor, it executes the relevant content of the method for rearranging recommended data as described above, or executes the relevant content of the method for rearranging recommended data in the catering service application platform as described above.

[0097] Based on the above content, the present application also provides a computer storage medium for storing data generated by a network platform and a program for processing the data generated by the network platform; When the program is read and executed by the processor, it executes the relevant content of the method for rearranging the recommended data as described above, or executes the relevant content of the method for rearranging the recommended data in the catering service application platform as described above.

[0098] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0099] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0100] Memory may include non-permanent storage in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0101] 1. Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can be implemented using any method or technology to store information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (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, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0102] 2. Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0103] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for rearranging recommended data, characterized in that: include: Get the advertisement recommendation type sorting sequence of the advertisement recommendation object; According to the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, obtaining at least one candidate sequence of the advertisement recommendation object, including a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; Performing sequence evaluation on the candidate sequence to obtain an indicator score corresponding to the estimated indicator type of the candidate advertisement recommendation object in the candidate sequence; including: obtaining an estimated click-through rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain a candidate fusion feature; and obtaining an indicator score corresponding to the estimated indicator type based on the candidate fusion feature; Obtaining, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; The candidate sequence that meets the re-arrangement score requirement and is selected from the sequence score range is determined as the re-arrangement sequence of the advertisement recommendation type ranking sequence.

2. The method for rearranging recommended data according to claim 1, wherein: The method of obtaining the advertisement recommendation object in the advertisement recommendation type sorting sequence based on the preference sequence generation type condition, the position generation type condition, and the refined sorting generation type condition in the heuristic sequence generation condition, and at least one candidate sequence among a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the advertisement recommendation type sorting sequence generation type condition, includes: The first candidate sequence is determined by using the estimated click-through rate, estimated click-through conversion rate, payment index, and weight coefficient of the estimated click-through conversion rate of the advertisement recommendation objects in the advertisement recommendation type sorting sequence as preference sequence generation type conditions; Determine the second candidate sequence by using one or more of the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, the context data, the estimated click-through rate, and the estimated conversion rate as position generation type conditions; The advertisement recommendation type ranking sequence is used as the third candidate sequence determined according to the refined ranking type generation condition.

3. The method for rearranging recommended data according to claim 2, wherein: The determining of the second candidate sequence by using one or more of the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, the context data, the estimated click-through rate, and the estimated conversion rate as position-based generation type conditions includes: One or more of the advertisement recommendation type ranking sequence, natural recommendation type ranking sequence, user portrait, context data, estimated click-through rate, and estimated conversion rate are input as input data for position generation type conditions into the position estimation model; The advertising recommendation type sorting sequence, the natural recommendation type sorting sequence, the user portrait, and at least one feature data of the context data in the input data are processed by the embedding layer of the position estimation model to obtain feature vector data; The feature vector data of the advertising recommendation type sorting sequence and the feature vector data of the natural recommendation type sorting sequence are processed by a cross-attention mechanism through the natural sequence encoding module in the position estimation model to obtain the relationship features between the advertising recommendation type sorting sequence and the natural recommendation type sorting sequence; Passing the ranked sequence of advertisement recommendation types through the advertisement sequence encoding module in the position estimation model to extract correlation features within the advertisement sequence; Passing the user portrait and the context data through the portrait / context encoding layer in the position estimation model to extract user attribute features and context features; fusing and connecting the estimated click rate, the estimated conversion rate, the relationship feature, the association feature, and one or more of the user attribute feature and the context feature through the position-based estimation model to obtain a fused connection feature; Processing the fused connection features through the hierarchical extraction mechanism of the position-based prediction model to obtain the position-based matrix of the estimated click-through rate and the position-based matrix of the estimated conversion rate; Determine a position matrix of an estimated click-through conversion rate by combining the position matrix of the estimated click-through rate and the position matrix of the estimated conversion rate of the advertisement recommendation object; The second candidate sequence is determined according to the position matrix of the estimated click-through conversion rate.

4. The method for rearranging recommended data according to claim 3, wherein: The determining the second candidate sequence according to the position matrix of the estimated click-through conversion rate includes: sequentially placing the advertisement recommendation objects in the advertisement recommendation type sorting sequence into positions in the estimated click-through conversion rate position matrix to obtain position scores of the advertisement recommendation objects at the positions; The advertisement recommendation objects are sorted by their position scores, and the advertisement recommendation object with the highest score is selected from the sorting results and placed into the position candidate sequence; The position-based candidate sequence is determined as the second candidate sequence.

5. The method for rearranging recommended data according to claim 1, wherein: The obtaining of the estimated click rate, estimated conversion rate, natural recommendation sequence, and advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence includes: Obtaining, based on the log data of the advertisement recommendation type sorting sequence, the estimated click rate, the estimated conversion rate, the natural recommendation sequence, and the advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; The step of fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the advertisement recommendation sequence, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features includes: Obtaining internal relationship features of the advertisement recommendation sequences among the advertisement recommendation sequences through a self-attention mechanism; Obtaining internal relationship features of the natural recommendation sequences between the natural recommendation sequences through the self-attention mechanism; Obtaining the correlation characteristics between the natural recommendation sequence and the advertising recommendation sequence through a multi-head target attention mechanism; Fusing the internal relationship features of the advertisement recommendation sequence, the internal relationship features of the natural recommendation sequence, the association relationship features, and the features of the candidate sequence to obtain candidate fusion features; Obtaining the indicator score corresponding to the estimated indicator type based on the candidate fusion features includes: The candidate fusion features are passed through a multi-layer extraction network model to obtain the estimated exposure rate, estimated click rate, and estimated conversion rate of the candidate advertising recommendation objects in the candidate sequence.

6. The method for rearranging recommended data according to claim 1, wherein: Obtaining, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score, includes: Determining a click-through conversion rate of the candidate advertisement recommendation object based on the estimated conversion rate and the estimated click-through rate of the candidate advertisement recommendation object in the indicator score; The sequence score of the candidate sequence is determined based on the estimated exposure rate, the estimated click-through rate, and the payment data and the estimated click-through conversion rate of the candidate advertisement recommendation object.

7. The method for rearranging recommended data according to claim 6, characterized in that: The step of selecting the candidate sequence that meets the re-arrangement score requirement from the sequence score range and determining it as the re-arrangement sequence of the advertisement recommendation type ranking sequence comprises: sorting the candidate sequences according to the sequence scores; The candidate sequence with the largest sequence score in the selected sorting is determined as the rearranged sequence of the advertisement recommendation type sorting sequence.

8. A device for rearranging recommended data, characterized in that: include: A first acquiring unit is configured to acquire an advertisement recommendation type sorting sequence of advertisement recommendation objects; a second acquisition unit, configured to obtain, based on the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, the advertisement recommendation object in the advertisement recommendation type sorting sequence, and at least one candidate sequence selected from a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; The third acquisition unit is used to perform sequence evaluation on the candidate sequence to obtain the index score corresponding to the estimated index type of the candidate advertising recommendation object in the candidate sequence; it includes: an acquisition subunit, used to obtain the estimated click-through rate, estimated conversion rate, natural recommendation sequence, and advertising recommendation sequence corresponding to the advertising recommendation type sorting sequence; a fusion subunit, used to fuse the estimated click-through rate, estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and correlation relationship features between the advertising recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; a calculation subunit, used to obtain the index score corresponding to the estimated index type based on the candidate fusion features; a fourth obtaining unit, configured to obtain, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; A determining unit is configured to determine the candidate sequence selected from the sequence score range and meeting the re-arrangement score requirement as a re-arrangement sequence of the advertisement recommendation type sorting sequence.

9. A method for rearranging recommended data in a catering service application platform, characterized in that: include: Obtaining a ranking sequence of advertisement recommendation types based on advertisement recommendation objects provided by the catering service application platform; According to the preference sequence generation type condition, the position generation type condition, and the refined ranking generation type condition in the heuristic sequence generation condition, the advertisement recommendation object in the advertisement recommendation type sorting sequence is obtained, and at least one candidate sequence is selected from a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the position generation type condition, and a third candidate sequence corresponding to the refined ranking generation type condition; Performing sequence evaluation on the candidate sequence to obtain an indicator score corresponding to the estimated indicator type of the candidate advertisement recommendation object in the candidate sequence; including: obtaining an estimated click-through rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; fusing the estimated click-through rate, the estimated conversion rate, the internal relationship features of the natural recommendation sequence, the features of the candidate sequence, and the correlation relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain a candidate fusion feature; and obtaining an indicator score corresponding to the estimated indicator type based on the candidate fusion feature; Obtaining, according to the indicator score, a sequence score of the candidate sequence corresponding to the indicator score; The candidate sequence that meets the re-arrangement score requirement and is selected from the sequence score range is determined as the re-arrangement sequence of the advertisement recommendation type ranking sequence.

10. A computer program product, characterized in that include: A computer program, which, when executed by a processor, implements the method for rearranging recommendation data as described in any one of claims 1 to 7 above, or executes the method for rearranging recommendation data in a catering service application platform as described in claim 9 above.

11. An electronic device, characterized in that: include: processor; A memory for storing a program for processing data generated by an electronic device, wherein when the program is read and executed by the processor, the program executes the method for rearranging recommended data as described in any one of claims 1 to 7 above, or executes the method for rearranging recommended data in a catering service application platform as described in claim 9 above.

Images