Method and apparatus for rearranging recommended data, computer program product and electronic device

Abstract

The application discloses a rearrangement method and device of recommended data, a computer program product and an electronic device, and the method comprises the following steps: acquiring an advertisement recommendation type sequence of an advertisement recommendation object; obtaining at least one candidate sequence in a first candidate sequence, a second candidate sequence and a third candidate sequence according to a preference sequence generation type condition, a position-specific generation type condition and a fine rearrangement generation type condition in a heuristic sequence generation condition; 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; obtaining the index score corresponding to the estimated index type according to a candidate fusion feature; obtaining a sequence score of the candidate sequence according to the index score; and determining a candidate sequence meeting a rearrangement score requirement selected from a sequence score range as a rearranged sequence of the advertisement recommendation type sequence, so that the data explosion problem caused by a large number of sequence combinations in the rearrangement process is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application and artificial intelligence, in particular to a method and device for rearranging recommended data. The present application also relates to a method and device for rearranging recommended data in a catering service application platform, and a computer program product, an electronic device and a computer storage medium. BACKGROUND

[0002] In today's digital world, providing information to users through application software has become a common technology. Based on this, recommending appropriate information to different users to achieve personalized information recommendation has become an important means to improve user experience. Therefore, a recommendation system has become an important bridge connecting users and businesses.

[0003] The content service recommended by the recommendation system can include natural service and advertising service, and can also be understood as natural recommendation and advertising recommendation. Sorting recommended content is a key technical link connecting users and application service content, and the core goal is to accurately predict and sort in a large amount of recommended data according to user interest, context environment and business target.

[0004] Generally, the process of a recommendation system generally includes retrieval, coarse ranking, fine ranking, re-ranking, etc. The existence of the two stages of fine ranking and re-ranking is mainly to optimize the quality of the recommended results at different levels to ensure that the recommended content finally displayed to the user not only meets the user's interest but also achieves the business target. Thus, the ecological balance of the entire service system is achieved. SUMMARY

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

[0006] The present application provides a method for rearranging recommended data, comprising:

[0007] obtaining an advertising recommendation type sorting sequence of an advertising recommendation object;

[0008] According to the preference sequence generation type condition, the position division generation type condition and the fine ranking generation type condition in the heuristic sequence generation condition, at least one candidate sequence is obtained from the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position division generation type condition and the third candidate sequence corresponding to the fine ranking generation type condition based on the advertising recommendation object;

[0009] The sequence evaluation is performed 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; 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 ranking sequence; performing fusion processing on the estimated click-through rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association 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 according to the candidate fusion features;

[0010] According to the index score, a sequence score of the candidate sequence corresponding to the index score is obtained;

[0011] The candidate sequence that meets the rearrangement score requirement is selected from the sequence score range, and is determined as a reordering sequence of the advertisement recommendation type ranking sequence.

[0012] In some embodiments, the preference sequence generation type condition, the position generation type condition, and the fine arrangement generation type condition in the heuristic sequence generation condition are used to obtain the advertisement recommendation object in the advertisement recommendation type ranking sequence based on at least one of 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, including:

[0013] The estimated click-through rate, the estimated click-through conversion rate, the payment index, and the weight coefficient of the estimated click-through conversion rate of the advertisement recommendation object in the advertisement recommendation type ranking sequence are used as the preference sequence generation type condition to determine the first candidate sequence;

[0014] 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 are used as the position generation type condition to determine the second candidate sequence;

[0015] The advertisement recommendation type ranking sequence is used as the third candidate sequence determined according to the fine arrangement generation type condition.

[0016] In some embodiments, the 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 are used as the position generation type condition to determine the second candidate sequence, including:

[0017] The 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 rate, and the estimated conversion rate are taken as input data of a position estimation model.

[0018] At least one feature data in the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait, and the context data in the input data is processed by an embedding layer of the position estimation model to obtain feature vector data.

[0019] The feature vector data of the advertisement recommendation type ranking sequence and the feature vector data of the natural recommendation type ranking sequence are processed by a natural sequence encoding module in the position estimation model through a cross-attention mechanism to obtain a relationship feature between the advertisement recommendation type ranking sequence and the natural recommendation type ranking sequence.

[0020] The advertisement recommendation type ranking sequence is extracted by an advertisement sequence encoding module in the position estimation model to obtain an association feature in the advertisement sequence.

[0021] The user portrait and the context data are extracted by a portrait / context encoding layer in the position estimation model to obtain a user attribute feature and a context feature.

[0022] One or more features of the estimated click rate, the estimated conversion rate, the relationship feature, the association feature, and the user attribute feature and the context feature are fused by a fusion connection of the position estimation model to obtain a fusion connection feature.

[0023] The fusion connection feature is processed by a hierarchical extraction mechanism of the position estimation model to obtain a position matrix of the estimated click rate and a position matrix of the estimated conversion rate.

[0024] The position matrix of the estimated click rate and the position matrix of the estimated conversion rate of the advertisement recommendation object are combined to determine a position matrix of an estimated click conversion rate.

[0025] The second candidate sequence is determined according to the position matrix of the estimated click conversion rate.

[0026] In some embodiments, the second candidate sequence is determined according to the position matrix of the estimated click conversion rate, including:

[0027] The advertisement recommendation object in the advertisement recommendation type ranking sequence is sequentially placed in a position in the position matrix of the estimated click conversion rate to obtain a position score of the advertisement recommendation object at the position.

[0028] The advertisement recommendation object with the first score in the sorting result of the score sorting of the advertisement recommendation objects according to the position score is selected and placed into the position candidate sequence;

[0029] The position candidate sequence is determined as the second candidate sequence.

[0030] In some embodiments, the obtaining of 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 comprises:

[0031] The estimated click rate, the estimated conversion rate, the internal relationship feature of the advertisement recommendation sequence, the internal relationship feature of the natural recommendation sequence, the feature of the candidate sequence, and the association relationship feature between the advertisement recommendation sequence and the natural recommendation sequence are fused to obtain candidate fusion features.

[0032] The internal relationship feature of the advertisement recommendation sequence is obtained through a self-attention mechanism.

[0033] The internal relationship feature of the natural recommendation sequence is obtained through the self-attention mechanism.

[0034] The association relationship feature between the natural recommendation sequence and the advertisement recommendation sequence is obtained through a multi-head target attention mechanism.

[0035] The advertisement recommendation sequence internal relationship feature, the natural recommendation sequence internal relationship feature, the association relationship feature, and the feature of the candidate sequence are fused to obtain candidate fusion features.

[0036] The estimated index type corresponding index score is obtained according to the candidate fusion features.

[0037] The candidate fusion features are input into a multi-layer extraction network model to obtain the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the candidate advertisement recommendation object in the candidate sequence.

[0038] In some embodiments, the sequence score of the candidate sequence corresponding to the index score is obtained according to the index score.

[0039] The click conversion rate of the candidate advertisement recommendation object is determined according to the estimated conversion rate and the estimated click rate of the candidate advertisement recommendation object in the index score.

[0040] determine a sequence score of the candidate sequence according to the estimated exposure rate, the estimated click rate, and payment data and an estimated click conversion rate of the candidate advertisement recommendation object.

[0041] In some embodiments, the candidate sequence meeting the rearrangement score requirement selected from the sequence score range is determined as a rearrangement column of the advertisement recommendation type ranking sequence.

[0042] sequence according to the sequence score;

[0043] The candidate sequence with the largest sequence score in the selected sorting is determined as a rearrangement column of the advertisement recommendation type ranking sequence.

[0044] The application also provides a rearrangement device for recommending data, comprising:

[0045] A first acquisition unit is configured to acquire an advertisement recommendation type ranking sequence of an advertisement recommendation object.

[0046] A second acquisition unit is configured to obtain at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ranking sequence based on a first candidate sequence corresponding to a preference sequence generation type condition, a second candidate sequence corresponding to a position generation type condition, and a third candidate sequence corresponding to a fine arrangement generation type condition in the heuristic sequence generation conditions.

[0047] A third acquisition unit is configured to perform sequence evaluation on the candidate sequence to obtain an estimated index score corresponding to an index type of a candidate advertisement recommendation object in the candidate sequence. The third acquisition unit comprises: an acquisition subunit configured to acquire an estimated click rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; a fusion subunit configured to fuse the estimated click rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and a calculation subunit configured to obtain the estimated index score corresponding to the index type according to the candidate fusion features.

[0048] A fourth acquisition unit is configured to obtain a sequence score of the candidate sequence corresponding to the index score according to the index score.

[0049] A determination unit is configured to determine the candidate sequence meeting the rearrangement score requirement selected from the sequence score range as a rearrangement column of the advertisement recommendation type ranking sequence.

[0050] The application further provides a rearrangement method of recommended data in a catering service application platform, comprising:

[0051] obtaining an advertisement recommendation type sequence of an advertisement recommendation object based on the catering service application platform;

[0052] According to the preference sequence generation type condition, the sub-position generation type condition and the fine arrangement generation type condition in the heuristic sequence generation condition, at least one candidate sequence in the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the sub-position generation type condition and the third candidate sequence corresponding to the fine arrangement generation type condition is obtained, and the advertisement recommendation object in the advertisement recommendation type sequence is obtained;

[0053] The sequence evaluation is performed 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, comprising: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence and an advertisement recommendation sequence corresponding to the advertisement recommendation type sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence and association 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 according to the candidate fusion features;

[0054] According to the index score, a sequence score of the candidate sequence corresponding to the index score is obtained;

[0055] The candidate sequence selected from the sequence score range and satisfying the rearrangement score requirement is determined as a rearrangement sequence of the advertisement recommendation type sequence.

[0056] The application further provides a computer program product, comprising: a computer program, which, when executed by a processor, implements the rearrangement method of recommended data or the rearrangement method of recommended data in a catering service application platform.

[0057] The application further provides an electronic device, comprising:

[0058] a processor;

[0059] a memory for storing a program for processing data generated by the electronic device, wherein the program, when read and executed by the processor, implements the rearrangement method of recommended data or the rearrangement method of recommended data in a catering service application platform.

[0060] Compared with the prior art, the application has the following advantages:

[0061] This application provides a method for rearranging recommendation data. In the heuristic sequence generation stage, corresponding first, second, and third candidate sequences are generated according to the preference sequence generation type condition, the positional generation type condition, and the refined ranking generation type condition in the heuristic sequence generation conditions. This avoids the data explosion problem caused by massive sequence combinations during the rearrangement process. In the sequence evaluation stage, the estimated click-through rate, estimated conversion rate, the internal relationship features of the natural recommendation sequence, the features of the candidate sequences, and the correlation features between the advertising recommendation sequence and the natural recommendation sequence are fused to obtain candidate fusion features and estimated index type scores for the candidate fusion features. By introducing an exposure probability index score and weighting the exposure probability in the sequence score calculation stage, the sequence evaluation value is made closer to the true value. Furthermore, by capturing contextual information in the sequence evaluation stage, more accurate predictions are obtained, and the context in the subsequent display stage is ensured to be consistent with the context in the evaluation stage. Attached Figure Description

[0062] Figure 1 This is a flowchart of a method for rearranging recommendation data provided in this application.

[0063] Figure 2 This is a schematic diagram of an embodiment of the rearrangement service invocation in a method for rearranging recommendation data provided in this application.

[0064] Figure 3 This is a schematic diagram of the location prediction model structure in a method for rearranging recommendation data provided in this application.

[0065] Figure 4 This is a schematic diagram of a network structure embodiment of the recommendation data rearrangement method provided in this application.

[0066] Figure 5 This is a schematic diagram of a data rearrangement device provided in this application.

[0067] Figure 6 This is a flowchart of a method for rearranging recommendation data in a catering service application platform provided in this application.

[0068] Figure 7 This is a schematic diagram of the structure of a data rearrangement device for recommendation data in a catering service application platform provided in this application.

[0069] Figure 8 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0070] To enable those skilled in the art to better understand the technical solutions of this application, the application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. However, this application can be implemented in many other ways different from those described below. Therefore, based on the embodiments provided in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0071] It should be noted that the terms "first," "second," "third," etc., in the claims, specification, and drawings of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. Such data are interchangeable where appropriate so that the embodiments of this application described herein can be implemented in a sequence other than that shown or described in this application. Furthermore, the terms "comprising," "having," and their variations are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0072] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0073] It should be understood that in the embodiments of this 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 that B is determined solely based on A; B can also be determined based on A and / or other information.

[0074] Based on the aforementioned background technology, it is clear that some life service-related applications, such as shopping, food delivery, and video streaming apps, can provide users with personalized or tailored content. The accuracy of these recommendations must satisfy both user experience and maintain a balance between the relationships between merchants, platforms, and users. This is essential for the application to operate and develop within an environment that caters to the needs of all parties. Therefore, the ranking of recommended content becomes a crucial technical means before it is displayed on electronic devices. In other words, ranking is a core issue of recommendation systems, aiming to provide users with a well-ordered queue of service information.

[0075] Ranking typically includes coarse ranking, fine ranking, and re-ranking. Coarse ranking primarily aims to quickly filter out suitable candidates from a massive candidate set, reducing the computational burden on subsequent steps. Fine ranking, building upon coarse ranking, further refines the candidate ranking to improve recommendation accuracy. Re-ranking considers a wider range of factors, such as merchant factors, user factors, and the application platform itself, requiring adjustments and optimizations to the fine ranking results. This ensures the recommended content displayed on the application platform achieves a balance from multiple perspectives and dimensions. Recommended content typically includes organic recommendations based on user preferences and advertising recommendations based on merchant and platform preferences.

[0076] However, in the existing technology, during the reordering stage, it is necessary to combine and mix the recommended advertising content into the naturally recommended content. This process generates a large number of combinations, which affects online processing performance and also causes time consumption.

[0077] In view of this, this application provides a method for rearranging recommendation data, which can avoid the problem of combined data explosion caused by a large amount of combined data due to rearrangement, and can meet the rights and interests of users, platforms and merchants.

[0078] like Figure 1 As shown, Figure 1 This is a flowchart of a method for rearranging recommendation data provided in this application, the method comprising:

[0079] Step S101: Obtain the sorted sequence of ad recommendation types for the ad recommendation object;

[0080] Step S102: Based on the preference sequence generation type condition, position-based generation type condition, and fine-ranking generation type condition in the heuristic sequence generation conditions, obtain at least one candidate sequence of the advertisement recommendation object based on the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position-based generation type condition, and the third candidate sequence corresponding to the fine-ranking generation type condition.

[0081] Step S103: 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; including: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association 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 according to the candidate fusion features;

[0082] Step S104: obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score;

[0083] Step S105: determining the candidate sequence selected from the sequence score range and satisfying the rearrangement score requirement as a rearrangement sequence of the advertisement recommendation type ranking sequence.

[0084] The above steps are described in detail as follows.

[0085] Regarding Step S101: obtaining an advertisement recommendation type ranking sequence of an advertisement recommendation object.

[0086] In this embodiment, the rearrangement is mainly for advertisement recommendation objects, and the rearrangement is performed based on the result of the above rearrangement based on the fine ranking sequence, and therefore, this step is based on the fine ranking result of the advertisement recommendation type ranking sequence of the advertisement recommendation object. This step is the basis for performing the rearrangement operation, such as Figure 2 as shown in Figure 2 is a schematic diagram of a rearrangement service calling embodiment in a recommendation data rearrangement method provided by the present application.

[0087] Based on an access request for an application service or a search request in the application service, a natural recommendation service and an advertisement recommendation service are respectively called to obtain a fine ranking sequence of a natural recommendation object and a fine ranking sequence of an advertisement recommendation object, and a mixed ranking service is called to trigger a rearrangement service request of the fine ranking sequence of the advertisement recommendation object through the mixed ranking service.

[0088] Therefore, in the specific implementation process of the step S101, the advertisement recommendation type sorting sequence of the advertisement recommendation object can be obtained by rearranging the service request, and the advertisement recommendation type sorting sequence can be a fine sorting sequence for the advertisement recommendation object. The trigger of the rearrangement of the service request can be based on the access request of the application service, or based on the search request of the search engine in the application service. In combination with the application scenario, it can be understood that the call of the rearrangement service can be triggered when the application service software is accessed, and the call of the rearrangement service can be triggered when 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 rearrangement service is applicable to a scenario where there is a sorting demand for recommendation data, for example, a food application scenario (including takeout, in-store service form), a shopping application scenario, and of course, a travel service application scenario.

[0089] Regarding step S102: According to the preference sequence generation type condition, the split position generation type condition and the fine arrangement generation type condition in the heuristic sequence generation condition, at least one of the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the split position generation type condition, and the third candidate sequence corresponding to the fine arrangement generation type condition of the advertisement recommendation object based on the preference sequence generation type condition is obtained.

[0090] Heuristic sequence generation refers to generating a sequence based on a heuristic algorithm. A heuristic algorithm is an algorithm that solves complex problems by finding empirical rules or heuristic knowledge. The algorithm explores the solution space of the problem to find a target solution. In this embodiment, three heuristic sequence generation conditions are provided to generate candidate sequences, thereby providing a basis for finding a target candidate sequence. The three heuristic sequence generation conditions can include: a preference sequence generation type condition, a split position generation type condition and a fine arrangement generation type condition. At least one of the three heuristic sequence generation conditions can be used in the specific implementation process. The specific implementation process of this step can include:

[0091] Step S102-1: The predicted click-through rate (pCTR: Predicted Click-Through Rate) of the advertisement recommendation object in the advertisement recommendation type sorting sequence, the predicted click-through and conversion rate (pCTCVR: Predicted Click-Through&Conversion Rate), the payment index (BID: advertisement bid), and the weight coefficient (ctcvr_weight) of the predicted click-through and conversion rate are taken as the preference sequence generation type condition to determine the first candidate sequence.

[0092] In this embodiment, the determination or generation of the first candidate sequence can be realized by using the following formula:

[0093]

[0094] wherein, a, b represent preference side super parameters; pCTR is the estimated click rate; bid is the advertisement bid; pCTCVR is the estimated click conversion rate; w is the weight coefficient of the estimated click conversion rate. By configuring a, b, different preferred first candidate sequences can be generated. w can also be configured with different weight coefficients.

[0095] Step S102-2: 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 rate, and the estimated conversion rate are taken as the position generation type condition to determine the second candidate sequence.

[0096] In this embodiment, the determination or generation of the second candidate sequence can adopt the following manner:

[0097] Step S102-21: 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 rate, and the estimated conversion rate are taken as the input data of the position generation type condition and input into the position estimation model; as shown in Figure 3 Figure 3 is a schematic diagram of a position estimation model structure in a recommendation data rearrangement method provided by the present application. The estimated click rate, the estimated conversion rate (pCTR / pCVR, pCTR: Predicted Click-Through Rate, estimated click rate, i.e. the probability of a user clicking after seeing an advertisement; pCVR: Predicted Conversion Rate, estimated conversion rate, i.e. the probability of a user completing a target behavior after clicking an advertisement), user portrait (Uer Proflie), context data (Context, which can be understood as the context of the fine ranking sequence), advertisement recommendation type ranking sequence (Ad Seq, which can also be referred to as advertisement sequence), natural recommendation type ranking sequence (Na Seq, which can also be referred to as natural sequence), which can be obtained through user behavior data, advertisement log data, etc. One or more of these data are taken as input data and input into the position estimation model to obtain the output second candidate sequence. The position estimation model can be constructed based on a list-wise sequence estimation model.

[0098] When the input data is input into the position estimation model, the corresponding processing process includes:

[0099] ​Step S102-22: At least one feature data in the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence, the user portrait and the context data in the input data is processed by an embedding layer of the sub-position estimation model to obtain feature vector data.

[0100] Step S102-23: The feature vector data of the advertisement recommendation type ranking sequence and the feature vector data of the natural recommendation type ranking sequence are processed by a natural sequence encoding module in the sub-position estimation model through cross attention mechanism to obtain relationship features between the advertisement recommendation type ranking sequence and the natural recommendation type ranking sequence. In this embodiment, the natural recommendation type ranking sequence includes an identification of a natural recommendation object, such as a commodity identification or a merchant identification, and related side information (side info), and features between the natural recommendation sequence and the advertisement recommendation sequence are extracted by a natural sequence encoding module (Na Encoder). In this embodiment, the Na Encoder module mainly includes a network including a multi-head target-attention mechanism, which can also be understood as a cross attention mechanism (Cross-Attention). The multi-head target-attention mechanism refers to an attention mechanism for dynamically calculating the importance of each item in a behavior sequence according to a target item when processing a user behavior sequence in a recommendation system. The Cross-Attention is an attention mechanism for establishing a connection between two different sequences. In fact, the multi-head target-attention mechanism can be regarded as an application form of the Cross-Attention. The Na Encoder module learns the related information between the advertisement and the natural recommendation sequence through the multi-head target-attention mechanism, that is, captures the influence of the natural recommendation sequence on the advertisement recommendation sequence, so that the relationship features between the advertisement recommendation type ranking sequence and the natural recommendation type ranking sequence can also be extracted by the Na Encoder module.

[0101] Step S102-24: The advertisement recommendation type ranking sequence is input into an advertisement sequence encoding module (Ad Encoder) in the sub-position estimation model to extract relevant features in the advertisement sequence. The relevant data of the advertisement recommendation type ranking sequence can include an item_id of the advertisement (an identification of an advertisement recommendation object, such as a product identification or a merchant identification, etc.), and relevant side information (sideinfo), which can be extracted by the advertisement sequence encoding module. The Ad Encoder module mainly includes a self-attention (self-attention mechanism) network to learn the relevant information between the internal sequences of the advertisement.

[0102] Step S102-25: The user portrait and the context data are input into a portrait / context encoding layer in the sub-position estimation model to extract user attribute features and context features. The context data can be the context of the advertisement recommendation type ranking sequence, such as time data, query data, and different types of context data. The user portrait can include attribute data of the user. The features involved in the context data and the attribute features involved in the user portrait can be extracted by the user and context encoding module.

[0103] Step S102-26: One or more features of the estimated click rate, the estimated conversion rate, the relationship features, the relevant features, and the user attribute features and context features are input into the fusion connection of the sub-position estimation model to obtain fusion connection features. In this embodiment, one of the pCTR / pCVR sub-position estimation model input data in the fine ranking stage is separately extracted and directly applied to the Concat&Fusion layer (concatenation and fusion layer) to strengthen the attention of the sub-position estimation model to the estimated click rate and the estimated conversion rate (pCTR / pCVR) features, improve the comprehensiveness of the data in the re-ranking stage, and provide a basis for the accuracy and adaptability of the subsequent re-ranking result. In this embodiment, the estimated click rate and the estimated conversion rate (pCTR / pCVR) can be input into the sub-position estimation model as a group of data, 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 group of data, because the user's click and conversion behaviors are positively correlated, therefore, introducing the estimated conversion rate into the estimation of the click rate when performing sub-position estimation on the sequence can have a positive guiding effect. In addition, by taking the estimated click rate and the estimated conversion rate (pCTR / pCVR) as a group of data, the consistency of the fine ranking link and the re-ranking link can be ensured. Of course, in other scenarios, the estimated click rate and the estimated conversion rate (pCTR / pCVR) can also be input as independent data.

[0104] Step S102-27: The fusion connection feature is processed by a hierarchical extraction mechanism of the sub-position estimation model to obtain a sub-position matrix of the estimated click rate and a sub-position matrix of the estimated conversion rate. The hierarchical extraction mechanism can be based on a PLE expert network, and shared features and task-specific features are gradually separated through multi-layer extraction. The PLE expert network can include two CGC networks, and each CGC network can include 3 specific-experts for CTR tasks and 3 specific-experts for CVR tasks. Two-way data corresponding to CTR tasks and CVR tasks are output, that is, the sub-position matrix of the estimated click rate and the sub-position matrix of the estimated conversion rate.

[0105] Step S102-28: The sub-position matrix of the estimated click rate and the sub-position matrix of the estimated conversion rate of the advertisement recommendation object are combined to determine a sub-position matrix of the estimated click conversion rate. The sub-position matrix of the estimated click rate and the sub-position matrix of the estimated conversion rate of the advertisement recommendation object are integrated through a fully connected layer, and then an estimated position result of pCTCVR (pCTRxpCVR) is obtained through an activation function (sigmoid). As shown in FIG. 8, slot1 represents the estimated value of position 1. Figure 3

[0106] In this embodiment, the sub-position estimation model can be constructed based on a List-wise network, list_size represents the length of the candidate sequence, and position_num represents the number of positions to be estimated, and then the dimension of the sub-position matrix can be represented as list_sizexposition_num. When calculating the loss, the sub-position matrix of the estimated click conversion rate and the sub-position matrix of the estimated click rate need to be maintained. The calculation formula of the loss is as follows:

[0107]

[0108] Wherein, y represents a sample label label (0 for a negative sample value of non-click or non-conversion or 1 for a positive sample value of click or conversion); i is the i-th advertisement; k is the k-th position; is the estimated value of the click rate or the conversion rate of the i-th advertisement at the k-th position, represents the calculation of cross entropy; is an indicator function, which takes 1 if the condition is true, and 0 otherwise, that is, whether the i-th advertisement is exposed at the k-th position, 1 for exposure and 0 for non-exposure.

[0109] By calculating the loss of each position of the sub-position matrix of the estimated click rate, a loss matrix and an estimated exposure sub-position matrix (​ ) Calculate Hadamard product, get loss matrix. It can be understood that the unexposed loss is masked using the estimated exposure sub-position matrix. In this embodiment, the loss of the exposure position is used.

[0110] The ESMM multi-task network respectively calculates the cross-entropy of CTR (Click-Through Rate: click rate) and CTCVR (Click-Through & Conversion Rate: click conversion rate), and the final loss is weighted by λ1 and λ2, as shown in the following formula:

[0111]

[0112] wherein, represents the click rate cross-entropy, represents the weight coefficient of the click rate cross-entropy; represents the click conversion rate cross-entropy, represents the weight coefficient of the click conversion rate cross-entropy.

[0113] Step S102-29: determining the second candidate sequence according to the estimated click conversion rate sub-position matrix. In this embodiment, the specific implementation process of this step can include:

[0114] Step S102-291: sequentially placing the advertisement recommendation objects in the advertisement recommendation type ranking sequence into the positions in the estimated click conversion rate sub-position matrix to obtain the position scores of the advertisement recommendation objects at the positions; in this embodiment, the position score can be calculated by the following formula:

[0115]

[0116]

[0117] wherein, is the reordered column click rate, is the advertisement bid, is the click conversion rate weight coefficient, is the estimated click rate.

[0118] Step S102-292: selecting the advertisement recommendation object with the first score in the sorting result of the score sorting of the advertisement recommendation objects according to the position score and placing it into the sub-position candidate sequence;

[0119] Step S102-293: determining the sub-position candidate sequence as the second candidate sequence.

[0120] Step S102-3: The advertisement recommendation type sorting sequence is determined as the third candidate sequence according to the fine sorting generation type condition. In this step, the fine sorting sequence of the obtained advertisement recommendation type sorting sequence can be directly taken as the third candidate sequence.

[0121] In this embodiment, three candidate types are obtained respectively based on three generation type conditions. The three candidate sequences corresponding to the three candidate types are taken as input data of the sequence evaluation model for sequence evaluation, that is, the step S103 is performed. In some other embodiments, one or more of the three generation type conditions can be optionally used to obtain the corresponding candidate sequence, and the three generation type conditions are not limited to being performed simultaneously, and the three generation type conditions are not limited to being used. For example, the coarse sorting generation condition can also be included, that is, the coarse sorting sequence for the advertisement recommendation object is directly taken as another candidate sequence.

[0122] Regarding step S103: The candidate sequence is evaluated to obtain the estimated index type corresponding index score of the candidate advertisement recommendation object in the candidate sequence.

[0123] The sequence evaluation in the step S103 can be processed by using a sequence evaluation model. Considering the influence of the mutual relationship between the natural recommendation sequence and the advertisement recommendation sequence on the reordering, the natural recommendation sorting and the advertisement recommendation sorting are split into two independent links for corresponding processing. From the modeling perspective of the sequence evaluation model, the natural recommendation and the advertisement recommendation are split into two sequences for separate modeling.

[0124] In the process of processing the advertisement recommendation sequence, the advertisement recommendation sequence is taken as one of the sequence evaluation model input feature data, and the self-attention mechanism is used to obtain the advertisement sequence internal relationship feature data between the advertisement recommendation sequences.

[0125] In the process of processing the natural recommendation sequence, the self-attention mechanism is used to obtain the natural sequence internal relationship feature between the natural recommendation sequences, and the natural sequence internal relationship feature is taken as one of the feature data. In addition, the Multi-Head Target Attention mechanism is used to process the relationship between the advertisement recommendation sequence and the natural recommendation sequence, to obtain the association relationship feature between the natural recommendation sequence and the advertisement recommendation sequence, and the association relationship feature is also taken as one of the feature data.

[0126] The pCTR and pCVR features of the advertisement recommendation sequence are taken as one of the input data of the sequence evaluation model, and are fused with the association relationship features and the internal relationship features to obtain fused features. For example, the pCTR values in the advertisement recommendation sequence are [0.3, 0.2, 0.1], the pCTR values in the natural recommendation sequence are [0.5, 0.05, 0.15, 0.35], and the pCTR values are sorted in ascending order as [0.05, 0.1, 0.15, 0.2, 0.3, 0.35, 0.5], wherein the bit positions corresponding to the advertisement recommendation sequence values are [4, 3, 1] (including 0), and the bit positions are taken as features and input into the sequence evaluation model.

[0127] Of course, the input data of the sequence evaluation model can also include user portraits, context data, and the like, and can also include other feature data in different application scenarios, for example, in the instant delivery scenario, the delivery distance, delivery price and the like.

[0128] The PLE Layer network layer (expert network layer) is used to obtain the estimated exposure rate, the estimated click rate and the estimated conversion rate. In this embodiment, the estimated exposure rate, the estimated click rate and the estimated conversion rate are weighted to the loss value in the sequence evaluation model training, so that the score obtained by the subsequent sequence evaluation model when evaluating the candidate sequence is more accurate.

[0129] Therefore, the specific implementation process of the step S103 can include:

[0130] 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 are obtained according to the log data of the advertisement recommendation type sorting sequence.

[0131] 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 are obtained according to the log data of the advertisement recommendation type sorting sequence.

[0132] The estimated click 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 association relationship features between the advertisement recommendation sequence and the natural recommendation sequence are fused to obtain candidate fused features.

[0133] The advertisement recommendation sequence internal relationship features between the advertisement recommendation sequences are obtained through the self-attention mechanism.

[0134] The natural recommendation sequence internal relationship features between the natural recommendation sequences are obtained through the self-attention mechanism.

[0135] An association relationship feature between the natural recommendation sequence and the advertisement recommendation sequence is obtained through a multi-head target attention mechanism.

[0136] The advertisement recommendation sequence internal relationship feature, the natural recommendation sequence internal relationship feature, the association relationship feature, and the feature of the candidate sequence are fused to obtain a candidate fusion feature.

[0137] The candidate fusion feature is obtained according to the candidate fusion feature, and the index score corresponding to the index type is obtained.

[0138] The candidate fusion feature is input into a multi-layer extraction network model to obtain an estimated exposure rate, an estimated click rate, and an estimated conversion rate of a candidate advertisement recommendation object in the candidate sequence.

[0139] Of course, the following can also be included:

[0140] According to the log data of the advertisement recommendation type ranking sequence, user portrait and context data are obtained, and the feature data of the user portrait, the feature data of the context, the advertisement recommendation sequence internal relationship feature, the natural recommendation sequence internal relationship feature, the association relationship feature, and the feature of the candidate sequence are fused to obtain a candidate fusion feature.

[0141] Regarding step S104, a sequence score of the candidate sequence corresponding to the index score is obtained according to the index score. In this step, the sequence score can be obtained using the following formula:

[0142]

[0143] wherein, represents an estimated exposure rate of the kth advertisement of the ith candidate sequence; represents an estimated click rate of the kth advertisement of the ith candidate sequence; represents a bid price of the kth advertisement of the ith candidate sequence; represents an estimated click conversion rate of the kth advertisement of the ith candidate sequence, and w represents a weight coefficient of the estimated click conversion rate.

[0144] Regarding step S105, the candidate sequence that meets the rearrangement score requirement is selected from the sequence score range, and is determined as a reordered sequence of the advertisement recommendation type ranking sequence.

[0145] The step S105 sorts the candidate sequences according to the sequence scores of the candidate sequences obtained in the step S104. The sorting can be in descending order or in ascending order. Regardless of the sorting type, the sorting result of the sorting of the candidate sequences and the corresponding sequence scores are selected according to the highest score in the candidate sequences, that is, the rearrangement score requirement. That is, the candidate sequence corresponding to the highest score is taken as the rearrangement column of the advertisement recommendation type sorting sequence. Based on the above sequence score calculation formula, the candidate sequence corresponding to the maximum value (max) is taken as the rearrangement column according to the score (rankScore) of the candidate sequence.

[0146] The above is a description of the specific execution process of the rearrangement method of the recommendation data provided by the application. In combination with Figure 4 , the Figure 4 is a schematic diagram of a network structure embodiment of the implementation principle of the rearrangement method of the recommendation data provided by the application. It can be known that the rearrangement method of the recommendation data provided by the application can be implemented based on the network structure shown in Figure 4 . The network structure can include 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, an expert layer, and the like. Based on the network structure in this embodiment and the above content, the implementation principle of the rearrangement method of the recommendation data in the application is described in outline.

[0147] The advertisement recommendation type sorting sequence, that is, the advertisement precision sorting column, is input into the heuristic sequence generation model. Through the sequence generation module in the heuristic sequence generation model, the first candidate sequence, the second candidate sequence, and the third candidate sequence are respectively generated according to the preference sequence generation type condition, the position generation type condition, and the precision generation type condition. The data involved in the preference sequence generation type condition includes the w weight, that is, the ctcvr_weight coefficient weight, the Bid index, the ctr index, and the like. 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 exemplified 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 realized through a greedy algorithm. The precision generation type condition can directly take the precision sorting column of the advertisement recommendation as the candidate sequence.

[0148] After generating the corresponding candidate sequence based on the above three generation conditions, the candidate sequence, the pCTR and pCVR of the advertisement recommendation type ranking sequence (i.e., the existing advertisement fine ranking sequence), the advertisement recommendation type ranking sequence, the natural recommendation type ranking sequence (the existing natural fine ranking sequence), etc. are processed as input data of the sequence evaluation model, and the purpose is to obtain the values of the candidate sequence in three dimensions, i.e., the values of pexp (estimated exposure rate), pctr (estimated click rate) and pcvr (estimated conversion rate). That is, the values of pexp, pctr and pcvr corresponding to the first candidate sequence, the values of pexp, pctr and pcvr corresponding to the second candidate sequence, and the values of pexp, pctr and pcvr corresponding to the third candidate sequence. Therefore, the sequence evaluation model includes modeling based on natural-advertisement sequence and multi-objective modeling based on exposure probability. Thus, in the processing process, not only the internal feature relationship within the natural recommendation type ranking sequence and the advertisement recommendation type ranking sequence can be focused on, but also the relationship features between the advertisement recommendation type ranking sequence and the advertisement recommendation type ranking sequence. By introducing the exposure probability and weighting pexp, pctr and pcvr, the subsequent sequence score calculation is more reasonable, because the exposure of the recommendation data will affect the user's behavior, therefore, the introduction of the exposure probability and the weight can avoid the deviation of the sequence score calculation.

[0149] Finally, the pexp, pctr and pcvr corresponding to the candidate sequence are calculated by the sequence selection module in the heuristic sequence generation sequence to obtain the score corresponding to the candidate sequence, and the candidate sequence with the highest score in the ranking is selected as the re-ranking sequence of the advertisement recommendation type ranking sequence and output.

[0150] In this embodiment, the position estimation model and the sequence evaluation model have similar network structures, and the difference is in the input data and the output data. The position information output by the position estimation model is used to obtain the second candidate sequence. For example, the position pCTR and pCVR of the advertisement recommendation object are obtained, all the advertisement recommendation objects are put into a preselected queue, the first position is selected to traverse and calculate the rankscore (formula in step S102) of all the advertisement recommendations, and the rankscore is sorted to select the candidate result queue that meets the score requirement and is removed from the preselected queue. The above operation is repeated for the next position until the preselected queue is empty, and the candidate sequence in the candidate result queue is the second candidate sequence, i.e., the position candidate sequence. For details, refer to the content of step S102 above, which will not be repeated here.

[0151] In multi-objective modeling based on exposure probability, the training process can involve selecting exposed samples as positive samples and unexposed samples as negative samples. Multi-task modeling combined with ESMM is used to predict CTR, CVR, and exp probabilities. The loss function is calculated using a weighted approach, as shown in the following formula:

[0152]

[0153] in, Indicates the estimated exposure loss ( The weighting coefficients of ) Indicates the loss of exposure click-through rate ( The weighting coefficients of ) Indicates the loss in exposure-click-conversion rate ( The weighting coefficients are: 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; expctr represents the estimated exposure click-through conversion rate, which is the product of the estimated exposure rate, the estimated click-through rate, and the estimated conversion rate.

[0154] The above is a summary description of the implementation principle of a recommendation data rearrangement method combined with a network model provided in this application. As can be seen from the above, in at least one embodiment of the recommendation data rearrangement method provided in this application, the required candidate sequences can be generated according to heuristic conditions during the heuristic sequence generation stage, thereby avoiding the data explosion problem caused by massive sequence combinations during the rearrangement process; the sequence evaluation stage can introduce exposure probability and weight it in the sequence selection score calculation stage, thereby making the sequence evaluation value closer to the true value. Furthermore, by capturing contextual information during the sequence evaluation stage, a more accurate prediction is obtained, and the context in the subsequent display stage is consistent with the context in the evaluation stage.

[0155] The above is a detailed description of an embodiment of a method for rearranging recommendation data provided in this application. Corresponding to the aforementioned embodiment of a method for rearranging recommendation data, this application also discloses an embodiment of a device for rearranging recommendation data. Please refer to [link / reference]. Figure 5 Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0156] like Figure 5 As shown, Figure 5 This is a schematic diagram of a data rearrangement device provided in this application, the device comprising:

[0157] The first acquisition unit 501 is used to acquire the sort sequence of the ad recommendation type of the ad recommendation object;

[0158] The second obtaining unit 502 is configured to obtain, according to the preference sequence generation type condition, the sub-position sequence generation type condition, and the fine sequence generation type condition in the heuristic sequence generation condition, at least one candidate sequence from among a first candidate sequence corresponding to the preference sequence generation type condition, a second candidate sequence corresponding to the sub-position sequence generation type condition, and a third candidate sequence corresponding to the fine sequence generation type condition based on the advertisement recommendation object.

[0159] The third obtaining unit 503 is configured to perform 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. The third obtaining unit includes: an obtaining subunit configured to obtain an estimated click rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; a fusion subunit configured to fuse and process the estimated click rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and a calculation subunit configured to obtain the index score corresponding to the estimated index type according to the candidate fusion features.

[0160] The fourth obtaining unit 504 is configured to obtain a sequence score of the candidate sequence corresponding to the index score according to the index score.

[0161] The determining unit 505 is configured to determine, as a reordering sequence of the advertisement recommendation type ranking sequence, the candidate sequence that meets the reordering score requirement and is selected from the sequence score range.

[0162] The specific content of the first obtaining unit 501 can refer to the above step S101, and will not be described in detail here.

[0163] The specific implementation process of the second obtaining unit 502 can include a first determining subunit, a second determining subunit, and a third determining subunit.

[0164] The first determining subunit is configured to determine the first candidate sequence by taking the estimated click rate, the estimated click conversion rate, the payment index, and the weight coefficient of the estimated click conversion rate of the advertisement recommendation object in the advertisement recommendation type ranking sequence as the preference sequence generation type condition.

[0165] The second determining subunit is configured to determine the second candidate sequence by taking 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 rate, and the estimated conversion rate as the sub-position sequence generation type condition.

[0166] The third determining subunit is configured to determine the third candidate sequence as the advertisement recommendation type sequence according to the fine arrangement generation type condition.

[0167] The second determining subunit can implement the following process:

[0168] The input subunit is configured to input one or more of the advertisement recommendation type sequence, the natural recommendation type sequence, the user portrait, the context data, the estimated click rate, and the estimated conversion rate as input data of the position-specific generation type condition to the position-specific estimation model.

[0169] The processing subunit is configured to obtain feature vector data by processing at least one feature data in the advertisement recommendation type sequence, the natural recommendation type sequence, the user portrait, and the context data in the input data through an embedding layer of the position-specific estimation model.

[0170] The first feature acquisition subunit is configured to obtain relationship features between the advertisement recommendation type sequence and the natural recommendation type sequence by performing cross-attention mechanism processing on the feature vector data of the advertisement recommendation type sequence and the feature vector data of the natural recommendation type sequence through a natural sequence encoding module in the position-specific estimation model.

[0171] The second feature acquisition subunit is configured to extract correlation features within an advertisement sequence by inputting the advertisement recommendation type sequence to an advertisement sequence encoding module in the position-specific estimation model.

[0172] The third feature acquisition subunit is configured to extract user attribute features and context features by inputting the user portrait and the context data to a portrait / context encoding layer in the position-specific estimation model.

[0173] The fusion subunit is configured to obtain fusion connection features by fusing one or more of the estimated click rate, the estimated conversion rate, the relationship features, the correlation features, and the user attribute features and context features through a fusion connection of the position-specific estimation model.

[0174] The first position-specific determining subunit is configured to obtain a position-specific matrix of the estimated click rate and a position-specific matrix of the estimated conversion rate by processing the fusion connection features through a hierarchical extraction mechanism of the position-specific estimation model.

[0175] The second position-specific determining subunit is configured to determine a position-specific matrix of the estimated click conversion rate by merging the position-specific matrix of the estimated click rate and the position-specific matrix of the estimated conversion rate of the advertisement recommendation object.

[0176] A second candidate sequence determining subunit is configured to determine the second candidate sequence according to the position-specific matrix of the estimated click conversion rate.

[0177] The specific implementation process of the second candidate sequence determining subunit can include:

[0178] A position score obtaining subunit is configured to sequentially place an advertising recommendation object in the advertising recommendation type ranking sequence into a position in the position-specific matrix of the estimated click conversion rate to obtain a position score of the advertising recommendation object at the position.

[0179] A selection subunit is configured to select an advertising recommendation object with the first score in a ranking result of the advertising recommendation objects ranked according to the position scores and place the advertising recommendation object into a position-specific candidate sequence.

[0180] A sequence determining subunit is configured to determine the position-specific candidate sequence as the second candidate sequence.

[0181] The obtaining subunit in the third obtaining unit 503 is specifically configured to obtain the estimated click rate, the estimated conversion rate, the natural recommendation sequence, and the advertising recommendation sequence corresponding to the advertising recommendation type ranking sequence according to log data of the advertising recommendation type ranking sequence. The fusion subunit is specifically configured to obtain an advertising recommendation sequence internal relationship feature between the advertising recommendation sequences by a self-attention mechanism, obtain a natural recommendation sequence internal relationship feature between the natural recommendation sequences by the self-attention mechanism, and obtain an association relationship feature between the natural recommendation sequences and the advertising recommendation sequences by a multi-head target attention mechanism. The fusion subunit is further configured to fuse the advertising recommendation sequence internal relationship feature, the natural recommendation sequence internal relationship feature, the association relationship feature, and a feature of the candidate sequence to obtain a candidate fusion feature. The calculation subunit is specifically configured to obtain the estimated exposure rate, the estimated click rate, and the estimated conversion rate of a candidate advertising recommendation object in the candidate sequence by a multi-layer extraction network model from the candidate fusion feature.

[0182] The specific implementation process of the fourth obtaining unit 504 can include: a first determining subunit configured to determine a click conversion rate of the candidate advertising recommendation object according to the estimated conversion rate and the estimated click rate of the candidate advertising recommendation object in the index score; and a second determining subunit configured to determine a sequence score of the candidate sequence according to the estimated exposure rate, the estimated click rate, payment data of the candidate advertising recommendation object, and the estimated click conversion rate.

[0183] The determining unit 505 specifically implements the process, including: an ordering subunit, configured to order the candidate sequences according to the sequence scores; and a determining subunit, configured to determine the candidate sequence with the maximum sequence score in the ordering as the reordering column of the advertisement recommendation type ordering sequence.

[0184] The above is a summary description of an embodiment of the reordering device for recommended data provided in the present application. For the specific implementation process of the device, refer to the related content of the above method embodiments.

[0185] Based on the above, the present application further provides a reordering method for recommended data in a catering service application platform, as shown in Figure 6 Figure 6 is a flowchart of a reordering method for recommended data in a catering service application platform provided in the present application. The method is mainly described by taking a catering service application platform or application software as an example, that is, the description of the reordering implementation of recommended data is in the application scenario of catering service. The catering service can be instant service such as takeout and retail. Of course, it is not limited to this application scenario. The method includes:

[0186] Step S601: obtaining an advertisement recommendation type ordering sequence of an advertisement recommendation object provided based on a catering service application platform;

[0187] Step S602: obtaining at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ordering sequence based on a preference sequence generation type condition corresponding to a first candidate sequence, a position generation type condition corresponding to a second candidate sequence, and a fine arrangement generation type condition corresponding to a third candidate sequence according to the preference sequence generation condition in the heuristic sequence generation condition;

[0188] Step S603: performing sequence evaluation on the candidate sequences to obtain an index score corresponding to an estimated index type of the candidate advertisement recommendation object in the candidate sequences; including: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type ordering sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association 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 according to the candidate fusion features;

[0189] Step S604: obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score;

[0190] ​Step S605: determining the candidate sequence selected from the sequence score range and satisfying the rearrangement score requirement as a rearrangement column of the advertisement recommendation type ordering sequence.

[0191] The specific content of steps S601-S605 can refer to the content of steps S101-S105, which will not be described in detail here.

[0192] Correspondingly, the application also provides a rearrangement device for recommendation data in a catering service application platform, as shown in Figure 7 Figure 7 is a structural schematic diagram of a rearrangement device for recommendation data in a catering service application platform provided by the application; the device comprises:

[0193] The first acquisition unit 701 is configured to acquire an advertisement recommendation type ordering sequence of an advertisement recommendation object provided based on a catering service application platform;

[0194] The second acquisition unit 702 is configured to obtain at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ordering sequence based on a preference sequence generation type condition, a position generation type condition, and a fine arrangement generation type condition in the heuristic sequence generation condition, the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position generation type condition, and the third candidate sequence corresponding to the fine arrangement generation type condition;

[0195] The third acquisition unit 703 is configured to perform 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; the third acquisition unit 703 comprises: an acquisition subunit configured to acquire an estimated click rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type ordering sequence; a fusion subunit configured to fuse the estimated click rate, the estimated conversion rate, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and a calculation subunit configured to obtain the index score corresponding to the estimated index type according to the candidate fusion features;

[0196] The fourth acquisition unit 704 is configured to obtain a sequence score of the candidate sequence corresponding to the index score according to the index score.

[0197] The determination unit 705 is configured to determine the candidate sequence selected from the sequence score range and satisfying the rearrangement score requirement as a rearrangement column of the advertisement recommendation type ordering sequence.

[0198] ​The above technical solutions of the present application can be applied to transactions, distribution services of instant e-commerce platforms, such as Taobao flash shopping, Taofxian, Eleme takeout and retail, etc.

[0199] Based on the above, the present application further provides a computer program product, comprising: a computer program, which, when executed by a processor, implements the related content of the above-mentioned recommendation data rearrangement method, or executes the related content of the above-mentioned recommendation data rearrangement method in the catering service application platform.

[0200] Based on the above, the present application further provides an electronic device, such as Figure 8 As shown in the figure, Figure 8 The present application provides a structural schematic diagram of an electronic device, which comprises:

[0201] a processor 801;

[0202] a memory 802 for storing a program for processing data generated by the electronic device, wherein the program, when read and executed by the processor, executes the related content of the above-mentioned recommendation data rearrangement method, or executes the related content of the above-mentioned recommendation data rearrangement method in the catering service application platform.

[0203] Based on the above, the present application further provides a computer storage medium for storing network platform generated data and a program for processing the network platform generated data;

[0204] The program, when read and executed by the processor, executes the related content of the above-mentioned recommendation data rearrangement method, or executes the related content of the above-mentioned recommendation data rearrangement method in the catering service application platform.

[0205] 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 for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal.

[0206] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0207] The memory can include non-persistent memory in computer readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer readable media.

[0208] 1. Computer-readable media includes permanent and non-permanent, removable and non-removable media can be implemented by 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 memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition herein, computer-readable media does not include transitory media, such as modulated data signals and carriers.

[0209] 2. Those skilled in the art should understand that the embodiments of the present application can be provided as a method, system or computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0210] Although the present application is disclosed with reference to the preferred embodiments above, it is not intended to limit the present application, and any person skilled in the art can 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 should be defined by the scope defined by the claims of the present application.

Claims

1. A method of recommending a rearrangement of data, characterized by, The method comprises the following steps: obtaining an advertisement recommendation type ranking sequence of an advertisement recommendation object; generating a preference sequence generation type condition, a position division generation type condition and a fine arrangement generation type condition according to heuristic sequence generation conditions, to obtain at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ranking sequence, the candidate sequence being based on the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position division generation type condition, and the third candidate sequence corresponding to the fine arrangement generation type condition; 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, comprising: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the advertisement recommendation sequence, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and obtaining a sequence score of the candidate sequence corresponding to the index score according to the candidate fusion features; obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score; selecting the candidate sequence that meets the rearrangement score requirement from the sequence score range to determine the rearranged sequence of the advertisement recommendation type ranking sequence.

2. The method of reordering recommended data according to claim 1, wherein, The method comprises the following steps: obtaining an advertisement recommendation type ranking sequence of an advertisement recommendation object; generating a preference sequence generation type condition, a position division generation type condition and a fine arrangement generation type condition according to heuristic sequence generation conditions, to obtain at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ranking sequence, the candidate sequence being based on the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position division generation type condition, and the third candidate sequence corresponding to the fine arrangement generation type condition; 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, comprising: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the advertisement recommendation sequence, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and obtaining a sequence score of the candidate sequence corresponding to the index score according to the candidate fusion features; 3. The method of rearranging recommended data according to claim 2, wherein, obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score; selecting the candidate sequence that meets the rearrangement score requirement from the sequence score range to determine the rearranged sequence of the advertisement recommendation type ranking sequence. The method comprises the following steps: obtaining an advertisement recommendation type ranking sequence of an advertisement recommendation object; generating a preference sequence generation type condition, a position division generation type condition and a fine arrangement generation type condition according to heuristic sequence generation conditions, to obtain at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ranking sequence, the candidate sequence being based on the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position division generation type condition, and the third candidate sequence corresponding to the fine arrangement generation type condition; 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, comprising: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the advertisement recommendation sequence, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and obtaining a sequence score of the candidate sequence corresponding to the index score according to the candidate fusion features; obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score; selecting the candidate sequence that meets the rearrangement score requirement from the sequence score range to determine the rearranged sequence of the advertisement recommendation type ranking sequence. The method comprises the following steps: obtaining an advertisement recommendation type ranking sequence of an advertisement recommendation object; generating a preference sequence generation type condition, a position division generation type condition and a fine arrangement generation type condition according to heuristic sequence generation conditions, to obtain at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ranking sequence, the candidate sequence being based on the first candidate sequence corresponding to the preference sequence generation type condition, the second candidate sequence corresponding to the position division generation type condition, and the third candidate sequence corresponding to the fine arrangement generation type condition; 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, comprising: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the advertisement recommendation sequence, internal relationship features of the natural recommendation sequence, features of the candidate sequence, and association relationship features between the advertisement recommendation sequence and the natural recommendation sequence to obtain candidate fusion features; and obtaining a sequence score of the candidate sequence corresponding to the index score according to the candidate fusion features; obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score; selecting the candidate sequence that meets the rearrangement score requirement from the sequence score range to determine the rearranged sequence of the advertisement recommendation type ranking sequence. The feature data in at least one of the advertisement recommendation type sequence, the natural recommendation type sequence, the user portrait and the context data in the input data is processed by an embedding layer of the sub-position estimation model to obtain feature vector data; The feature vector data of the advertisement recommendation type sequence and the feature vector data of the natural recommendation type sequence are processed by a natural sequence encoding module in the sub-position estimation model through a cross attention mechanism to obtain relationship features between the advertisement recommendation type sequence and the natural recommendation type sequence; The advertisement sequence encoding module in the sub-position estimation model is used to extract correlation features within the advertisement sequence from the advertisement recommendation type sequence; The user attribute features and context features are extracted from the user portrait and the context data by a portrait / context encoding layer in the sub-position estimation model; One or more features of the estimated click rate, the estimated conversion rate, the relationship features, the correlation features, and the user attribute features and context features are fused by the fusion connection of the sub-position estimation model to obtain fusion connection features; The fusion connection features are processed by a hierarchical extraction mechanism of the sub-position estimation model to obtain a sub-position matrix of the estimated click rate and a sub-position matrix of the estimated conversion rate; The sub-position matrix of the estimated click rate and the sub-position matrix of the estimated conversion rate of the advertisement recommendation object are combined to determine a sub-position matrix of the estimated click conversion rate; The second candidate sequence is determined according to the sub-position matrix of the estimated click conversion rate.

4. The method of rearranging recommended data according to claim 3, wherein, The second candidate sequence is determined according to the sub-position matrix of the estimated click conversion rate, including: The advertisement recommendation objects in the advertisement recommendation type sequence are sequentially placed in the positions in the sub-position matrix of the estimated click conversion rate to obtain position scores of the advertisement recommendation objects at the positions; An advertisement recommendation object with the first score in a sorting result of the advertisement recommendation objects sorted according to the position scores is selected and placed in a sub-position candidate sequence; The sub-position candidate sequence is determined as the second candidate sequence.

5. The method of rearranging recommended data of claim 1, wherein, The estimated click rate, the estimated conversion rate, the natural recommendation sequence, and the advertisement recommendation sequence corresponding to the advertisement recommendation type sequence are obtained, including: The estimated click rate, the estimated conversion rate, the natural recommendation sequence, and the advertisement recommendation sequence corresponding to the advertisement recommendation type sequence are obtained according to log data of the advertisement recommendation type sequence; The estimated click 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 are fused and processed to obtain candidate fusion features, including: The advertisement recommendation sequence internal relationship features between the advertisement recommendation sequences are obtained through a self-attention mechanism; The natural recommendation sequence internal relationship features between the natural recommendation sequences are obtained through the self-attention mechanism; The association relationship feature between the natural recommendation sequence and the advertisement recommendation sequence is obtained through a multi-head target attention mechanism; The estimated click rate, the estimated conversion rate, the internal relationship feature of the advertisement recommendation sequence, the internal relationship feature of the natural recommendation sequence, the association relationship feature, and the feature of the candidate sequence are fused to obtain a candidate fusion feature; The estimated index type corresponding index score is obtained according to the candidate fusion feature, including: The estimated exposure rate, the estimated click rate, and the estimated conversion rate of the candidate advertisement recommendation object in the candidate sequence are obtained through a multi-layer extraction network model based on the candidate fusion feature.

6. The method of reordering recommended data according to claim 5, wherein, The sequence score of the candidate sequence corresponding to the index score is obtained according to the index score, including: The estimated click conversion rate of the candidate advertisement recommendation object is determined according to the estimated conversion rate and the estimated click rate of the candidate advertisement recommendation object in the index score; The sequence score of the candidate sequence is determined according to the estimated exposure rate, the estimated click rate, the payment data of the candidate advertisement recommendation object, and the estimated click conversion rate.

7. The method of reordering recommended data according to claim 6, wherein, The candidate sequence that meets the rearrangement score requirement is selected from the sequence score range to determine the reorder column of the advertisement recommendation type sorting sequence, including: The candidate sequence is sorted according to the sequence score; The candidate sequence with the maximum sequence score in the selected sorting is determined as the reorder column of the advertisement recommendation type sorting sequence.

8. A device for recommending rearrangement of data, characterized by Including: A first acquisition unit is configured to acquire an advertisement recommendation type sorting sequence of an advertisement recommendation object; A second acquisition unit is configured to obtain at least one candidate sequence of the advertisement recommendation object based on a preference sequence generation type condition, a position generation type condition, and a precision arrangement generation type condition in a heuristic sequence generation condition, from the advertisement recommendation type sorting sequence, 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 precision arrangement generation type condition; A third acquisition unit is configured to perform sequence evaluation on the candidate sequence to obtain an estimated index type corresponding index score of a candidate advertisement recommendation object in the candidate sequence, including: an acquisition subunit configured to acquire an estimated click rate, an estimated conversion rate, a natural recommendation sequence, and an advertisement recommendation sequence corresponding to the advertisement recommendation type sorting sequence; a fusion subunit configured to fuse and process the estimated click rate, the estimated conversion rate, an internal relationship feature of the advertisement recommendation sequence, an internal relationship feature of the natural recommendation sequence, a feature of the candidate sequence, and an association relationship feature between the advertisement recommendation sequence and the natural recommendation sequence to obtain a candidate fusion feature; and a calculation subunit configured to obtain the estimated index type corresponding index score according to the candidate fusion feature; A fourth acquisition unit is configured to obtain a sequence score of the candidate sequence corresponding to the index score according to the index score. The determining unit is configured to determine the candidate sequence meeting the rearrangement score requirement selected from the sequence score range as a rearrangement column of the advertisement recommendation type ranking sequence.

9. A method for rearranging recommendation data in a catering service application platform, characterized in that, The method comprises: obtaining an advertisement recommendation type ranking sequence of an advertisement recommendation object provided by a catering service application platform; obtaining at least one candidate sequence of the advertisement recommendation object in the advertisement recommendation type ranking sequence based on a preference sequence generation type condition, a position generation type condition and a fine arrangement generation type condition in the heuristic sequence generation condition; 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; the sequence evaluation comprises: obtaining an estimated click rate, an estimated conversion rate, a natural recommendation sequence and an advertisement recommendation sequence corresponding to the advertisement recommendation type ranking sequence; performing fusion processing on the estimated click rate, the estimated conversion rate, internal relationship features of the advertisement recommendation sequence, internal relationship features of the natural recommendation sequence, features of the candidate sequence and association 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 according to the candidate fusion features; obtaining a sequence score of the candidate sequence corresponding to the index score according to the index score; determining the candidate sequence meeting the rearrangement score requirement selected from the sequence score range as a rearrangement column of the advertisement recommendation type ranking sequence.

10. A computer program product, characterised in that, The method comprises: a computer program, when executed by a processor, implements the rearrangement method of recommendation data according to any one of claims 1-7 or the rearrangement method of recommendation data in a catering service application platform according to claim 9.

11. An electronic device, comprising: The method comprises: a processor; a memory configured to store a program for processing data generated by an electronic device, the program, when read and executed by the processor, implements the rearrangement method of recommendation data according to any one of claims 1-7 or the rearrangement method of recommendation data in a catering service application platform according to claim 9.

Images