Store ranking, display method and device, equipment and storage medium

Abstract

The present disclosure provides a store ranking, display method, device, equipment and storage medium, the store ranking method comprises: obtaining the interest of a user, recalling a plurality of candidate stores for the user; after obtaining the store enhancement feature of the candidate store by the store scoring model, determining the probability information of the user canceling the interest in the candidate store; wherein the store enhancement feature is obtained after the information of the user's preference and the interest suitable for canceling the interest in the store feature of the candidate store is enhanced; according to the probability information of each candidate store, the plurality of candidate stores are ranked.

Description

Technical Field

[0001] This disclosure relates to the field of Internet technology, and in particular to store sorting, display methods, devices, equipment and storage media. Background Technology

[0002] In local services and related fields, offering user benefits is a crucial means of user acquisition and engagement, managing users throughout their entire lifecycle. Therefore, many related internet product designs include pages featuring benefits and store recommendations. These pages typically display specific benefits at the top, with a list of stores below for redemption. Currently, the store recommendation scheme for benefit distribution recalls multiple candidate stores, sorts them based on a certain strategy, and then recommends them to users. However, poor store sorting can negatively impact user redemption of benefits. Therefore, it is necessary to provide a store sorting method to improve the accuracy of store recommendations and facilitate user redemption of benefits. Summary of the Invention

[0003] To overcome the problems existing in the related technologies, this disclosure provides a method, apparatus, device and storage medium for sorting and displaying stores.

[0004] According to a first aspect of the embodiments of this specification, a store sorting method is provided, the method comprising:

[0005] To obtain the user's rights, multiple candidate stores were recalled for the user;

[0006] After obtaining the enhanced features of the candidate stores from the store scoring model, the probability information of the user redeeming the benefits in the candidate stores is determined; wherein, the enhanced features are obtained by enhancing the representation of the information in the store features of the candidate stores that the user prefers and is suitable for redeeming the benefits;

[0007] The candidate stores are sorted according to their probability information.

[0008] Optionally, the store enhancement features are obtained in the following way:

[0009] Obtain the feature sequence of the historical popular redemption store set corresponding to the benefit, and obtain the feature sequence of the user's historical preferred store set;

[0010] By utilizing the feature sequences of the historical popular redemption store set and the feature sequences of the historical preferred store set, target features that represent user preferences and are suitable for redeeming the benefits are obtained;

[0011] Obtain the store features of the candidate stores, and enhance the representation of the information related to the target features in the store features of the candidate stores to obtain enhanced store features.

[0012] Optionally, the target feature includes: a first target feature, a second target feature, or a third target feature; wherein,

[0013] The first target feature is obtained by enhancing the representation of information related to the feature sequence of the historical preferred store set in the feature sequence of the historical popular redemption store set;

[0014] The second target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preference store set;

[0015] The third target feature is obtained by fusing the first target feature and the second target feature.

[0016] Optionally, the store rating model includes:

[0017] A first Transformer structure is used to enhance the representation of information related to the feature sequence of the historically popular redemption store set in the feature sequence of the historically preferred store set, and output a first target feature; or...

[0018] A second Transformer structure is used to enhance the representation of information related to the feature sequences of the historical popular redemption stores in the feature sequences of the historical preference store set, and output a second target feature; or...

[0019] The system comprises a first Transformer structure, a second Transformer structure, and a connection layer. The connection layer is connected to the output of the first Transformer structure and to the output of the second Transformer structure. The connection layer is used to concatenate the first target feature output by the first Transformer structure and the second target feature output by the second Transformer structure to output the third target feature.

[0020] Optionally, the feature sequence of the historical popular redemption store set is obtained in the following way:

[0021] Obtain the store characteristics of each store in the historical popular redemption store set;

[0022] The information related to the store characteristics of the candidate stores is enhanced in the store characteristics of each store.

[0023] The enhanced store characteristics of each store are combined into a feature sequence.

[0024] Optionally, the historical popular redemption store set corresponding to the benefit includes: the user's location and the historical popular redemption store set corresponding to the benefit.

[0025] Optionally, the method further includes:

[0026] Obtain historical redemption data within a set time range, and group and statistically analyze it according to geographical location and benefit threshold. After obtaining multiple sets of historical popular redemption stores, obtain the feature sequence of each set of historical popular redemption stores and store it.

[0027] The feature sequence of obtaining the historical popular redemption store set corresponding to the rights includes:

[0028] From the feature sequences of each set of historical popular redemption stores stored in the database, obtain the feature sequence corresponding to the user's location and the benefit threshold.

[0029] Optionally, after obtaining the enhanced features of the candidate stores from the store scoring model, determining the probability information of the user redeeming the benefits in each of the candidate stores includes:

[0030] After obtaining the benefit characteristics of the benefits and the enhanced characteristics of the candidate stores from the store scoring model, the probability information of the user redeeming the benefits in each of the candidate stores is determined;

[0031] The rights and benefits features include basic features that characterize the basic information of the rights and benefits, and / or combined features obtained by combining the basic features with any of the following features: user features of the user, scene features of the current scene, store features of the candidate store, or user behavior features towards the candidate store.

[0032] Optionally, obtaining user rights includes:

[0033] Get the user's currently claimed benefits.

[0034] Optionally, the method further includes:

[0035] Extract the set number of candidate stores with the highest probability information from the sorted candidate stores and send them to the client.

[0036] According to a second aspect of the embodiments of this specification, a store display method is provided, the method comprising:

[0037] In response to a store display request, the system retrieves multiple sorted candidate stores sent by the server. These candidate stores are obtained by the server based on the user's benefits and are then recalled for that user. After obtaining enhanced features of the candidate stores using a store scoring model, the system determines the probability information of the user redeeming the benefits at those candidate stores. Based on the probability information of each candidate store, the system sorts the multiple candidate stores. The enhanced features are obtained by strengthening the representation of information in the candidate stores that reflects the user's preferences and is suitable for redeeming the benefits.

[0038] Display the multiple sorted candidate stores.

[0039] Optionally, the store enhancement features are obtained in the following way:

[0040] Obtain the feature sequence of the historical popular redemption store set corresponding to the benefit, and obtain the feature sequence of the user's historical preferred store set;

[0041] By utilizing the feature sequences of the historical popular redemption store set and the feature sequences of the historical preferred store set, target features that represent user preferences and are suitable for redeeming the benefits are obtained;

[0042] Obtain the store features of the candidate stores, and enhance the representation of the information related to the target features in the store features of the candidate stores to obtain enhanced store features.

[0043] Optionally, the target feature includes: a first target feature, a second target feature, or a third target feature; wherein,

[0044] The first target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preferred store set;

[0045] The second target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preference store set;

[0046] The third target feature is obtained by fusing the first target feature and the second target feature.

[0047] Optionally, the store rating model includes:

[0048] A first Transformer structure is used to enhance the representation of information related to the feature sequence of the historically popular redemption store set in the feature sequence of the historically preferred store set, and output a first target feature; or...

[0049] A second Transformer structure is used to enhance the representation of information related to the feature sequences of the historical popular redemption stores in the feature sequences of the historical preference store set, and output a second target feature; or...

[0050] The system comprises a first Transformer structure, a second Transformer structure, and a connection layer. The connection layer is connected to the output of the first Transformer structure and to the output of the second Transformer structure. The connection layer is used to concatenate the first target feature output by the first Transformer structure and the second target feature output by the second Transformer structure to output the third target feature.

[0051] Optionally, the feature sequence of the historical popular redemption store set is obtained in the following way:

[0052] Obtain the store characteristics of each store in the historical popular redemption store set;

[0053] The information related to the store characteristics of the candidate stores is enhanced in the store characteristics of each store.

[0054] The enhanced store characteristics of each store are combined into a feature sequence.

[0055] Optionally, the historical popular redemption store set corresponding to the benefit includes: the user's location and the historical popular redemption store set corresponding to the benefit.

[0056] Optionally, obtaining the feature sequence of the historical popular redemption store set corresponding to the rights includes:

[0057] Obtain historical redemption data within a set time range, and group and statistically analyze it according to geographical location and benefit threshold. After obtaining multiple sets of historical popular redemption stores, obtain the feature sequence of each set of historical popular redemption stores and store it.

[0058] From the feature sequences of each set of historical popular redemption stores stored in the database, obtain the feature sequence corresponding to the user's location and the benefit threshold.

[0059] Optionally, after obtaining the enhanced features of the candidate stores from the store scoring model, determining the probability information of the user redeeming the benefits in each of the candidate stores includes:

[0060] After obtaining the benefit characteristics of the benefits and the enhanced characteristics of the candidate stores from the store scoring model, the probability information of the user redeeming the benefits in each of the candidate stores is determined;

[0061] The rights and benefits features include basic features that characterize the basic information of the rights and benefits, and / or combined features obtained by combining the basic features with any of the following features: user features of the user, scene features of the current scene, store features of the candidate store, or user behavior features towards the candidate store.

[0062] Optionally, obtaining user rights includes:

[0063] Get the user's currently claimed benefits.

[0064] Optionally, the method further includes:

[0065] Extract the set number of candidate stores with the highest probability information from the sorted candidate stores and send them to the client.

[0066] According to a third aspect of the embodiments of this specification, a store sorting device is provided, the device comprising:

[0067] The recall module is used to: obtain the user's rights and interests, and recall multiple candidate stores for the user;

[0068] The scoring module is used to: obtain the enhanced features of the candidate stores from the store scoring model, and then determine the probability information of the user redeeming the benefits in the candidate stores; wherein, the enhanced features are obtained by enhancing the representation of the information in the store features of the candidate stores that the user prefers and that is suitable for redeeming the benefits.

[0069] The sorting module is used to sort the multiple candidate stores according to the probability information of each candidate store.

[0070] According to a fourth aspect of the embodiments of this specification, a store display device is provided, the device comprising:

[0071] The acquisition module is configured to: respond to a store display request, acquire multiple sorted candidate stores sent by the server; the candidate stores are those retrieved by the server based on the user's benefits; after acquiring the enhanced features of the candidate stores using a store scoring model, determine the probability information of the user redeeming the benefits at the candidate stores; and sort the multiple candidate stores according to the probability information of each candidate store; wherein, the enhanced features are obtained by enhancing the representation of the information in the store features of the candidate stores that is preferred by the user and suitable for redeeming the benefits.

[0072] The display module is used to display the multiple sorted candidate stores.

[0073] According to a fifth aspect of the embodiments of this specification, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method embodiments described in the first or second aspect above.

[0074] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method embodiments described in the first or second aspect above.

[0075] The technical solutions provided in the embodiments of this specification may include the following beneficial effects:

[0076] In this embodiment, user benefits are obtained, and multiple candidate stores are recalled for the user. After obtaining the enhanced features of the candidate stores by the store scoring model, the probability information of the user redeeming the benefits in each of the candidate stores is determined. The enhanced features are obtained by strengthening the representation of the information in the store features of the candidate stores that the user prefers and that is suitable for redeeming the benefits, so that the model can score the stores based on the benefits information, associate the recommended stores with the benefits, and thus improve the benefits redemption rate.

[0077] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0078] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this specification and, together with the specification, serve to explain the principles of this disclosure.

[0079] Figure 1A This is a schematic diagram of a client-side rights distribution page illustrated in this specification according to an exemplary embodiment.

[0080] Figure 1B This is a schematic diagram illustrating a server and a client according to an exemplary embodiment of this specification.

[0081] Figure 2A This is a flowchart illustrating a store sorting method according to an exemplary embodiment of this specification.

[0082] Figure 2B This is a schematic diagram illustrating an application scenario of store sorting according to an exemplary embodiment.

[0083] Figure 2C This is a schematic diagram illustrating a store rating model according to an exemplary embodiment of this specification.

[0084] Figure 2D This is a schematic diagram illustrating an LBS rights activation sequence module according to an exemplary embodiment of this specification.

[0085] Figure 2E This is a schematic diagram illustrating an Attention structure according to an exemplary embodiment of this specification.

[0086] Figure 2F This is a schematic diagram illustrating a dual-sequence module according to an exemplary embodiment of this specification.

[0087] Figure 3 This is a flowchart illustrating a store display method according to an exemplary embodiment of this specification.

[0088] Figure 4 This is a hardware structure diagram of a computer device illustrated in this specification according to an exemplary embodiment.

[0089] Figure 5 This is a block diagram illustrating a layout device for a page module according to an exemplary embodiment of this specification.

[0090] Figure 6 This is a block diagram illustrating a page display device according to an exemplary embodiment of this specification. Detailed Implementation

[0091] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0092] The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this specification. The singular forms “a,” “the,” and “the” as used in this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0093] It should be understood that although the terms first, second, third, etc., may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this specification, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0094] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points shall be provided for users to choose to authorize or refuse.

[0095] In local services and related fields, coupon rewards are a crucial tool for attracting new users and boosting user activity, permeating the entire user lifecycle. Therefore, many related internet product designs feature pages dedicated to rewards and store recommendations. These pages typically display specific rewards at the top, with a list of stores below serving as the redemption mechanism. Thus, the store recommendations below the rewards directly impact redemption rates; the accuracy of the recommendations is directly related to the redemption rate. Therefore, optimizing the store recommendations below the reward pages is of paramount importance.

[0096] like Figure 1A The image shown is a schematic diagram of a client-side benefits distribution page according to an exemplary embodiment of this specification. The top of the page displays benefits with different thresholds and discounts, such as a red envelope with a discount of 4 yuan off for purchases over 35 yuan (35 yuan is the usage threshold and 4 yuan is the discount amount). The bottom of the page recommends a list of relevant stores that can redeem the benefits.

[0097] Benefits and privileges can influence the perceived value of products displayed to users, directly or indirectly affecting their personalized choices and purchasing intentions. For example, if a user has received a ¥5 discount coupon for purchases over ¥40, recommending stores with products priced above ¥40 might be more appropriate. Similarly, if a user has received a coupon for milk tea, recommending milk tea shops might be easier for them to redeem.

[0098] However, the relevant store ranking schemes usually only consider feature information such as users, stores, or spatiotemporal scenarios (time / space), mainly considering stores from the perspective of user preferences, but do not take into account the input of rights data and the expression of relevant rights information. Therefore, the correlation between the rights claimed by users and the recommended stores is poor, which makes it difficult for users to quickly redeem and convert their rights.

[0099] Based on this, the embodiments of this specification provide a store sorting scheme that can optimize the store list under the background of benefits, optimize store recommendations in combination with the benefits received by users, so that the recommended stores match the benefits received by users, the store recommendation accuracy is high, and it is convenient for users to quickly redeem and convert benefits.

[0100] Please see Figure 1B , Figure 1B This is a schematic diagram of a server and a client provided in an embodiment of this specification, including a server 100 and at least one client 200. For example, the client 200 can access the server 100 through the network to use the services provided by the server 100, including but not limited to goods delivery services, goods purchase services, reading services, audio and video playback services, or store search services, etc.

[0101] Server 100 can be a program installed in a background device to provide services to users. For example, such as... Figure 1B As shown, the backend device can be a server, which can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0102] Client 200 can be a program installed on a user's device to provide services to the user. Client 200 includes, but is not limited to, applications (APP), web pages, mini-programs, plugins, or components. Figure 1B As shown, user devices include, but are not limited to, smartphones, personal digital assistants, tablets, personal computers, laptops, wearable devices, virtual reality terminal devices, or augmented reality terminal devices.

[0103] like Figure 2A As shown, Figure 2A This is a flowchart illustrating a store sorting method according to an exemplary embodiment, comprising the following steps:

[0104] In step 202, the user's rights are obtained, and multiple candidate stores are recalled for the user.

[0105] In step 204, after obtaining the enhanced features of the candidate stores from the store scoring model, the probability information of the user redeeming the rights in the candidate stores is determined.

[0106] The enhanced store feature is obtained by enhancing the representation of the user preference and the information suitable for redeeming the rights in the store features of the candidate store.

[0107] In step 206, the candidate stores are sorted according to their probability information.

[0108] The store sorting method provided in the embodiments of this specification can be derived from... Figure 1B This embodiment does not impose any restrictions on either the server 100 or the client 200 for execution. Taking server 100 as an example, client 200 can detect a store display request (e.g., a user's client operation to access a store display page or a benefits redemption page, etc.) and send it to server 100. Server 100 executes steps 202 to 206 as described above, obtains the ranking of the candidate stores, and then sends it to client 200 so that client 200 can display the stores according to the ranking. Taking client execution as an example, client 200 can detect a user's store display request operation, determine that a store display request has been initiated, and execute steps 202 to 206 as described above to obtain the ranking of the candidate stores, and then display the stores according to the ranking.

[0109] There are multiple ways to acquire user benefits. These benefits can be those displayed on a page when the user browses the page; they can be benefits already claimed by the user or unclaimed benefits; they can also be benefits the user has previously claimed. There are various ways to acquire user benefits. For example, the client can provide a page for users to claim benefits and display an entry point to that page. Once the client detects that the user has triggered this entry point, it can confirm that the user has claimed the benefit, and then allow the user to enter the page and execute subsequent steps. For instance, the client can send a message to the server indicating that the user has claimed the benefit, and the server can then execute the aforementioned store sorting example.

[0110] In this embodiment, there are multiple ways to implement the multiple candidate stores for user recall. Optionally, a preset recall system can return a set of candidate stores including multiple candidate stores. The recall strategy can be flexibly determined according to actual needs, and this embodiment does not limit it. As an example, the candidate stores for recall can include stores where the currently claimed benefits can be redeemed, or they can be determined by combining time and / or geographical location, such as stores that are open during the current time period, stores within a set distance from the user's current location, etc., or predicted stores that match the user's preferences, or a combination of one or more of the above factors, etc.

[0111] For the multiple candidate stores identified during the recall, this embodiment utilizes a pre-defined store scoring model to predict the probability of a user redeeming benefits at a candidate store. For example, the store scoring model can be invoked once for each candidate store, scoring that candidate store (i.e., outputting the probability of a user redeeming benefits at the candidate store). After all candidate stores have received scores, they are sorted. The store scoring model will be explained below.

[0112] Based on the model's output objective, the training samples for the store rating model in this embodiment can be obtained through user behavior data related to stores. For example, samples can be constructed using store data exposed to users on the historical benefits page, where exposure refers to the store being displayed on the screen of the client's electronic device. The page where users claim benefits includes a module displaying stores, which shows multiple stores. User behaviors related to these exposed stores include clicking and purchasing. Based on this, the model can be constructed using user behavior logs from exposure, clicks, and purchases in the store recommendation module below the benefits page. For example, by first associating user exposure, click, and purchase logs, samples labeled with positive click labels and / or positive purchase labels can be generated for user clicks and purchases, respectively. Samples of negative labels that were not clicked after exposure and / or negative labels that were not purchased after exposure can also be constructed.

[0113] The information input into the model can be configured as needed. For example, it can include the following four aspects:

[0114] a) User information may include basic user profiles (such as profile features like gender, age, economic capacity, and spending power) and / or historical user statistics (such as the number of times a user logged into the client within a set historical time range, the number of times a user made transactions within a set historical time range, etc.).

[0115] b) Store information may include basic store attribute information (such as store rating, store category, store consumption level, etc.) and store statistical information (such as the number of transactions and exposures within a set historical time range).

[0116] c) Scene information, which may include time information (current time), spatial information (current user's geographical location), etc.

[0117] d) User store cross-information, which represents user behavior information about stores (such as whether the user has clicked on the store in history, whether the user has purchased from the store in history, whether the user has redeemed benefits from the store in history, whether the user has favorited the store in history, etc.).

[0118] e) Rights and benefits information may include user-related rights and benefits information, such as information on the rights and benefits that the user has received, including the threshold for the rights and benefits, the level of discounts on the rights and benefits, the type of rights and benefits, and the usage period of the rights and benefits.

[0119] Optionally, this information can be normalized or discretized accordingly, and based on this, the model can extract the corresponding features from the input information.

[0120] As an example, in order to enhance the expression of rights and benefits information, the probability information of the user redeeming the rights and benefits in each of the candidate stores can be determined by obtaining the rights and benefits characteristics and the enhanced store characteristics of the candidate stores from the store scoring model.

[0121] The rights and benefits features include basic features that characterize the basic information of the rights and benefits, and / or combined features obtained by combining the basic features with any of the following features: user features of the user, scene features of the current scene, store features of the candidate store, or user behavior features towards the candidate store.

[0122] In this embodiment, the model also incorporates benefit features during prediction. Besides the basic benefit features, all other features are combined with them, ensuring that each feature carries benefit information. This enhances the representation of benefit information, allowing the model to focus more on benefit information when scoring candidate stores and output more accurate probability information. The combination of basic features with other features can be understood as feature intersection. Features are represented by vectors, and the Cartesian product of two feature vectors can be used as the combined feature. For example, the Cartesian product of a feature vector representing a basic feature and a feature vector representing a user feature represents the combined feature of the basic and user features. Other combined features are obtained in the same way.

[0123] Regarding step 204, the store scoring model for candidate stores enhances the representation of user preferences and the information suitable for redeeming the rights in the store characteristics of the candidate stores, resulting in enhanced store characteristics. This enhances the expression of user preferences and rights information in the store characteristics of candidate stores, making the recommended stores suitable for redemption of rights in this case if they meet user preferences.

[0124] Enhancing the representation of features with specific information can be achieved in various ways. For example, the feature vector can be weighted according to the information that needs to be enhanced; the resulting weighted feature is the enhanced feature. The weights can be configured as needed. For instance, an attention mechanism can be used.

[0125] In some examples, the store enhancement features are obtained in the following ways:

[0126] Obtain the feature sequence of the historical popular redemption store set corresponding to the benefit, and obtain the feature sequence of the user's historical preferred store set;

[0127] By utilizing the feature sequences of the historical popular redemption store set and the feature sequences of the historical preferred store set, target features that represent user preferences and are suitable for redeeming the benefits are obtained;

[0128] Obtain the store features of the candidate stores, and enhance the representation of the information related to the target features in the store features of the candidate stores to obtain enhanced store features.

[0129] The historical popular redemption store set includes multiple historically popular redemption stores, and the feature sequence of the historical popular redemption store set includes the feature vectors of each store in the set. Each historical popular redemption store can be obtained based on the statistical data of multiple users' historical usage rights placed in the store within a set time period.

[0130] In this embodiment, historically popular redemption stores can be grouped based on the benefit threshold. Since these historically popular redemption stores correspond to the benefit threshold, the extracted redemption correlation features include information related to the benefit threshold, representing the characteristics of historically popular redemption stores under the benefit threshold. This makes the probability of candidate stores predicted by the model being redeemed related to the benefit threshold, and the final score of the candidate stores is also related to the benefit threshold, improving the accuracy of the scoring and making the benefit and recommended stores highly correlated.

[0131] In other examples, the data obtained can also include multiple historically popular redemption stores corresponding to the user's location and benefit threshold. Based on this, the solution in this embodiment can be applied to business scenarios related to geolocation. For example, in a local life service scenario, stores provide services based on geolocation. This embodiment utilizes multiple historically popular redemption stores corresponding to the user's location, and the redemption relevance features obtained include geolocation-related information, representing the characteristics of historically popular redemption stores in that geolocation. Therefore, this embodiment can express differentiated information about redeemable stores under different benefit thresholds and different regions, improving the accuracy of store scoring.

[0132] In some cases, the above feature vector sequence can be pre-stored so that it can be quickly retrieved when needed.

[0133] For example, the method in this embodiment may further include:

[0134] Obtain historical redemption data within a set time range, and group and statistically analyze it according to geographical location and benefit threshold. After obtaining multiple sets of historical popular redemption stores, obtain the feature sequence of each set of historical popular redemption stores and store it.

[0135] The feature sequence of obtaining the historical popular redemption store set corresponding to the rights includes:

[0136] From the feature sequences of each set of historical popular redemption stores stored in the database, obtain the feature sequence corresponding to the user's location and the benefit threshold.

[0137] In this embodiment, the server can pre-store multiple sets of feature vector sequences corresponding to historically popular redemption stores. Specifically, historical redemption data within a set time range is acquired, grouped and statistically analyzed based on geographical location and benefit thresholds to obtain multiple sets of historically popular redemption stores. The store characteristics of each store within each set are obtained and combined to obtain and store the feature sequence of each set. Each set of feature vector sequences can be stored using geographical location and / or benefit thresholds as keywords. Geographical location can be represented by a certain geographical area or by GeoHash (an address encoding method). The benefit threshold can include a threshold value or a threshold range.

[0138] Based on this, when scoring online, the system can use the user's location and benefit threshold as keywords to quickly retrieve the feature vector sequence corresponding to the user's location and benefit threshold from multiple stored feature vector sequences of historically popular redemption stores.

[0139] In some examples, to enhance the representation of candidate store-related information in the feature sequence of the historical popular redemption store set, the feature sequence of the historical popular redemption store set can be obtained in the following way:

[0140] Obtain the store characteristics of each store in the historical popular redemption store set;

[0141] The information related to the store characteristics of the candidate stores is enhanced in the store characteristics of each store.

[0142] The enhanced store characteristics of each store are combined into a feature sequence.

[0143] As an example, enhanced representations can be achieved using attention mechanisms. Taking the calculation of the inner product as an example, the closer two objects are, the larger the inner product. The formula for the attention mechanism can be:

[0144]

[0145] Q refers to the query, K to the key, and V to the value. Attention can be described as mapping the query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors. The output is calculated as a weighted sum of the values, where the weight assigned to each value is calculated using a compatibility function between the query and the corresponding key.

[0146] In this embodiment, Q corresponds to the target store vector, and K and V are the feature vector sequences corresponding to historically popular redemption stores.

[0147] Attention essentially calculates a weight based on the query and key, then reassigns a weight to the value. During the calculation, the similarity between two vectors is first calculated based on K and Q. The softmax function compresses various input values ​​into the range of 0-1, thus revealing the strength of attention distributed across the values. Multiplying by V yields a new vector output containing the attention information.

[0148] The softmax function is used to calculate the similarity between the query and the key. In this embodiment, the scaled dot-product method is used to calculate the inner product of QK, where T represents the transpose of the matrix, and then multiplied by a scaling factor. Where d k This refers to the dimensions of query and key. Other functions can also be used in practice, but this embodiment does not limit this approach.

[0149] By using the above methods, the information representation related to candidate stores in the feature vector sequence corresponding to historically popular redemption stores can be activated, making the model output more accurate.

[0150] In this embodiment, the feature sequence of the historical popular redemption store set represents the store information suitable for redeeming the currently claimed benefits, and the feature sequence of the historical preferred store set represents the user's historical personalized preference information. Step 204 can combine the sequence representations of the two different dimensions to enhance the expression of benefits information.

[0151] In some examples, the target feature may include: a first target feature, a second target feature, or a third target feature; wherein,

[0152] The first target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preferred store set;

[0153] The second target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preference store set;

[0154] The third target feature is obtained by fusing the first target feature and the second target feature.

[0155] In this embodiment, information related to the feature sequences of the historical preferred store set can be extracted from the feature sequences of the historical popular redemption store set for representation. Alternatively, information related to the feature sequences of the historical preferred store set can be extracted from the feature sequences of the historical popular redemption store set for representation. Both methods can obtain store information describing user preferences and suitable for redeeming the benefits. Optionally, this embodiment also employs a third target feature that integrates the first and second target features, allowing the third target feature to combine two sequence representations from different dimensions and extract relevant information between the two sequences. The extraction of any of the above target features can be achieved in various ways in practical applications, such as through a feature extraction layer in a neural network.

[0156] In some examples, the store rating model includes any of the following structures:

[0157] (1) A first Transformer structure, wherein the first Transformer structure is used to enhance the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preferred store set, and output a first target feature; or,

[0158] (2) A second Transformer structure, which is used to enhance the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preference store set, and output a second target feature; or,

[0159] (3) A first Transformer structure, a second Transformer structure and a connection layer, wherein the connection layer is connected to the output of the first Transformer structure and to the output of the second Transformer structure, and the connection layer is used to concatenate the first target feature output by the first Transformer structure and the second target feature output by the second Transformer structure to output the third target feature.

[0160] The Transformer structure in this embodiment can be flexibly selected according to actual needs. For example, the Co-attention Transformer structure of the multimodal algorithm ViLBERT can be used, and this embodiment does not limit it. For the case of using a third target feature, this embodiment uses two Transformer structures and connects their outputs through a connection layer. The connection layer can fuse the two outputs to obtain the final feature. For example, the outputs of the two Transformer structures can be concatenated into a vector, and then processed with the candidate shops using an attention mechanism to obtain the final output.

[0161] The model structure of the store rating model in this embodiment can be configured as needed. For example, a multi-task model, such as the Entire Space Multi-Task Model (ESMM), can be used. The output of the model is pCTCVR (page Click-Through & Conversion Rate), which is the product of the click-through rate pCTR (page Click-Through Rate) and the conversion rate pCVR (page Conversion Rate), representing the probability that a user will place an order and redeem their benefits in the store.

[0162] In the above embodiments, after recalling n candidate stores, a store scoring model is invoked for each candidate store. Each invocation outputs a score for a candidate store. After obtaining the scores of the n candidates, these n candidate stores can be sorted. For example, higher scores indicate higher probabilities. The m candidate stores with the highest scores can be selected and sent to the client so that the client can display stores according to the sorting of the m candidate stores. Here, m is a set number that can be configured as needed. This embodiment does not limit this.

[0163] The following example illustrates this further. Taking the application of this example to the server side as an example, as a case study, the server side can be configured as follows: Figure 2B The system shown comprises multiple systems, each performing a different processing step. The overall architecture consists of two parts: an online phase and an offline phase. The online phase mainly comprises three components: a store ranking system, a benefits distribution system, and a model-based online scoring system. The online store ranking system first requests real-time benefits information from the benefits distribution system, then calls the model-based online scoring system to score and rank the stores, finally returning a list of stores. The offline phase mainly includes three parts: online log feedback processing, feature processing and sample generation, and model training. Finally, the offline-trained model is deployed to the online model-based online scoring system to provide online scoring services. Therefore, the following sections will describe the entire system separately for online and offline phases.

[0164] I. Offline Part:

[0165] 1. Log data collection: Data on online exposure, clicks, and purchase behavior of returning users, as well as real-time rights and benefits information of returning users.

[0166] 2. Sample Construction: The composition of the store sample is usually based on the user's exposure, clicks, and purchase behavior logs of stores in the store recommendation section below the benefits. First, the user exposure, click, and purchase logs are linked, and click positive labels and purchase positive labels are generated separately for user clicks and purchases.

[0167] 3. Feature extraction:

[0168] The following five features can be extracted:

[0169] a) User characteristics, including basic user profiles and historical user statistical characteristics;

[0170] b) Store characteristics: mainly including basic store attributes and statistical characteristics;

[0171] c) Scene characteristics: temporal and spatial characteristics, etc.;

[0172] d) User store cross-features (whether the user clicked on the store, whether the user purchased the product, whether the user added the product to their favorites, etc.).

[0173] e) Rights and benefits characteristics: mainly including user-related rights and benefits data.

[0174] This information is then normalized or discretized accordingly.

[0175] 4. Model building:

[0176] As an example, the entire model structure is as follows: Figure 2C As shown, the overall structure of the model adopts the ESMM multi-task model, which represents the equity features at the model feature level. This can be divided into three parts: a basic equity feature module, an LBS equity activation sequence module, and an equity dual sequence module. Each feature is fused through an all-feature embedding layer and then input to multiple fully connected layers (fc layers). The figure shows an example with six fully connected layers: three connected layers are used to output pCTR; the other three are used to output pCVR. The model's prediction result, pCTCVR, is used as the final output. The following sections describe these three modules in detail:

[0177] (1) Basic equity characteristics module;

[0178] This module is used to extract basic features and combined features that characterize the basic information of the rights.

[0179] ① The basic characteristics of the benefits include the benefit threshold, the amount of the discount, the category of the benefit (such as milk tea red envelopes), the type of benefit (such as basic red envelopes and expanded red envelopes), and the benefit claiming time. This benefit information is generated by directly calling the benefit distribution system. First, these basic benefit characteristics are directly used as input to the model.

[0180] ② Cross the basic features and other features (features such as users, stores or scenarios, see step 3 above) as model input.

[0181] (2) LBS rights activation sequence module;

[0182] Considering the spatiotemporal characteristics of local life service scenarios, this embodiment combines LBS (Location Based Services) with historical store best-selling information to construct a store sequence corresponding to different benefits under different LBS, expressing differentiated information on redeemable benefits (thresholds) and different regions.

[0183] ① Store LBS benefit activation sequences; based on the client's historical set time period (e.g., one month, 60 days, or 90 days), benefit redemption data, such as order data for consumption using benefits, can be grouped and statistically analyzed using GeoHash location codes and benefit redemption thresholds. This will generate a sequence of TOP popular redemption stores for different threshold benefits within each geographical range (e.g., the top 50 can be selected based on the total number of redeemed orders; the specific value can be configured as needed). Side information features of each popular redemption store, such as category, brand, and store tags, will be obtained and added to the sequence to serve as the LBS benefit activation sequence.

[0184] That is, this embodiment can store multiple sets of feature vector sequences of historical popular redemption stores (referred to as LBS benefit activation sequences in this embodiment), and the key of each sequence is: GeoHash location code + benefit threshold;

[0185] The key value is a sequence, which includes the feature vectors of N historically popular redemption stores. Each store can be understood as a node in the sequence, and each node corresponds to the feature vector of that store.

[0186] ② When scoring online, the model uses the GeoHash location code and the benefit threshold as keys to query the corresponding LBS benefit redemption sequence, which is the feature sequence of the aforementioned historical popular redemption store set (the representation has not been enhanced yet).

[0187] After obtaining the LBS benefits redemption sequence, the module in this embodiment can use a Target Attention structure to model the correlation between the benefits activation sequence and the target item, thereby further activating the information representation related to the target store in the LBS benefits redemption sequence. The Attention calculation in this embodiment uses an inner product approach; the closer the two stores are, the larger the inner product. See the structure below. Figure 2D , Figure 2D item t This is the feature vector for the target store (the target store during the training phase, i.e., the store with the label; during the online scoring phase, i.e., the candidate store), meaning one vector per store, and each vector is represented by many features. The LBS rights redemption sequence includes item1 to item2 in the diagram. N The target store's store vector is composed of N historically popular redemption stores, as shown in the figure. By determining the correlation between the target store's store vector and the store vectors of the N historically popular redemption stores, we can obtain N enhanced feature vectors of the N historically popular redemption stores. These feature vectors are then combined into a feature sequence, which is the activated LBS rights redemption sequence. In this embodiment, it can be called the LBS rights activation sequence.

[0188] When calculating the correlation between the target store's store vector and the store vector of each historically popular redemption store in the LBS benefits redemption sequence, as an example, the following attention mechanism can be used, with the calculation formula as follows:

[0189]

[0190] Where Q corresponds to the target store vector, and K and V are the store vectors in the LBS benefits redemption sequence. In this embodiment, the vector dimension is 64. Figure 2E The diagram shown is a schematic of the Attention structure, where the three Linear transformations represent vectors Q, K, and V, respectively, Scaled Dot-Product represents the inner product, and matmul represents the matrix multiplication operation.

[0191] The above embodiments can be used to characterize information related to candidate stores in the activated historical popular redemption store sequence.

[0192] (3) Equity Dual-Sequence Module:

[0193] The above embodiments introduced LBS rights activation sequences and verified the effectiveness of incorporating LBS rights activation sequence features. For example... Figure 2FAs shown, this embodiment uses the user's historical preferred stores to express the user's historical personalized preference information. To better combine the sequence representations of two different dimensions and enhance the expression of benefits information, this embodiment can use the Co-attention Transformer structure of the multimodal algorithm ViLBERT (a multimodal model for both vision and text) to extract relevant information between the two sequences, use one sequence to enhance the expression of the other sequence, and add the output to the main network.

[0194] Co-attention employs two Transformer structures, with the current sequence as Q and another sequence as K and V inputs respectively. The two sequences are then swapped to obtain two Transformer outputs. These two outputs are concatenated into a single vector, which is then processed with the target store's TargetAttention to obtain the final output.

[0195] The formula for calculating Co-attention is as follows:

[0196] H u =Transformer(Q u K c V c )

[0197] H c =Transformer(Q c K u V u )

[0198] Q target =H shop

[0199] K t =V t =concat(H u H c )

[0200] H T =Transformer(Q target K t V t )

[0201] Among them, H u and H c That is, the outputs of the two Transformer structures mentioned above, H u This indicates that information suitable for rights redemption from the user's historical preferred store sequence will be enhanced and represented. c This means that information in the LBS rights activation sequence that matches user preferences is enhanced and represented.

[0202] Q target and H shop That is, the target store; concat means to include H u and H c splicing; H T H represents the store characteristics of the target store. u and H c The information is enhanced and represented.

[0203] The Transformer is calculated as follows:

[0204]

[0205] H h =Multihead(Q,K,V)=Concat(head1,head2,…,head n W H

[0206] head i =Attention(QW i Q KW i K VW i V )

[0207]

[0208]

[0209] In this approach, Attention calculates the similarity between stores using the inner product. Q represents the query, K represents the key, and V represents the value.

[0210] Multihead attention, also known as multi-head attention, includes n heads and W heads. H The weight matrix is ​​represented by Multihead, which calculates it by linearly projecting Q, K, and V into multiple parallel spaces.

[0211] head i Let QW represent the i-th attention in multi-head attention. i Q KW i K VW i V Let W represent the matrices obtained by linearly transforming Q, K, and V using the weight matrix W, respectively.

[0212] LayerNorm represents the normalization layer;

[0213] FeedForward represents the feedforward layer.

[0214] In this embodiment, considering that the feature sequence of a user's historical preferred stores may not necessarily be suitable for rights redemption, the information suitable for rights redemption in the feature sequence of historical preferred stores is enhanced. The mutual enhancement of the two sequences is that one enhances the representation of the information suitable for rights redemption in the user's historical preferred store sequence, and the other enhances the representation of the information that matches the user's preferences in the LBS rights activation sequence.

[0215] The two enhanced feature sequences are concatenated to obtain a new feature sequence. This new feature sequence is then subjected to attention processing with the features of the candidate stores, so that the store features of the candidate stores that are related to user preferences and suitable for rights redemption are enhanced.

[0216] 5. Model Training and Deployment: After the model network is built in step 4 above, the model can be trained in a supervised manner based on features and labels. Optimization methods can include AdaGrad (Adaptive Gradient). Depending on the needs, the trained model can be stored on a distributed file system and then deployed to an online model scoring system.

[0217] II. Online Section:

[0218] 1. User Request: When a user enters the benefits store page, they initiate a request to the store ranking system.

[0219] 2. Obtaining Benefits: The store ranking system first calls the benefits distribution system to obtain the real-time benefits that users are currently receiving, including data such as the benefit threshold, discount amount, benefit category (e.g., milk tea coupons), benefit type (e.g., basic coupons, expanded coupons), and benefit collection time.

[0220] 3. Call the model for online scoring: The store ranking system processes the acquired rights data into the input format required by the model, then obtains the user's entry page request information (such as user_id) and the candidate store set recalled from the system (including multiple store ID sets). Based on this data, a specified format request is constructed as the input to the model online scoring system. The model online scoring system predicts and returns the predicted score for each store in real time.

[0221] 4. Return Recommendation Results: The store ranking system sorts stores in descending order based on the model's scoring results, then adds relevant store images, titles, and other information, and finally selects the top-scoring candidate stores to return to the user.

[0222] As can be seen from the above embodiments, this embodiment innovatively builds a real-time user benefits acquisition link in the store ranking system link. The acquired real-time benefits data is used as input to build a basic benefits feature module, an LBS benefits activation sequence module, and a benefits dual sequence module, which are added to the existing ESMM multi-task model network. This strengthens the expression of user benefits information and can better model the store ranking under benefits.

[0223] In practical applications, a user may claim multiple different benefits at once. The method of this embodiment can be flexibly executed as needed. For example, one or more benefits (e.g., the one with the greatest discount) can be selected to execute steps 202 to 206 above; or steps 202 to 206 above can be executed for all benefits. This embodiment does not limit this.

[0224] like Figure 3 The diagram shown is a flowchart illustrating a store display method according to an exemplary embodiment of this specification. This method can be applied to a client and includes the following steps:

[0225] In step 302, in response to the store display request, multiple sorted candidate stores sent by the server are obtained.

[0226] The candidate stores are multiple candidate stores recalled by the server for the user; after obtaining the enhanced features of the candidate stores by the store scoring model, the probability information of the user redeeming the benefits in each of the candidate stores is determined; the multiple candidate stores are ranked according to the probability information of the candidate stores; wherein, the enhanced features are obtained by enhancing the representation of the information in the store features of the candidate stores that the user prefers and is suitable for redeeming the benefits.

[0227] In step 304, the sorted candidate stores are displayed.

[0228] The above embodiments can be referred to the description of the foregoing embodiments, and will not be repeated here.

[0229] Corresponding to the aforementioned embodiments of the store sorting method / store display method, this specification also provides embodiments of the store sorting method / store display method and the computer equipment used thereon.

[0230] The embodiments of the store sorting / store display method apparatus described in this specification can be applied to computer devices, such as servers or terminal devices. The apparatus embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by its processor reading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective, such as... Figure 4 The diagram shown is a hardware structure diagram of a computer device containing the store sorting device / store display method device 431 in this specification. (Except for...) Figure 4 In addition to the processor 410, memory 430, network interface 420, and non-volatile memory 440 shown, the computer device in which the store sorting device / store display method device 431 in the embodiment is located may also include other hardware depending on the actual function of the computer device, which will not be described in detail here.

[0231] like Figure 5 As shown, Figure 5 This is a block diagram illustrating a store sorting device according to an exemplary embodiment of this specification, the device comprising:

[0232] Recall module 51 is used to: obtain the user's rights and interests, and recall multiple candidate stores for the user;

[0233] The scoring module 52 is used to: after obtaining the enhanced features of the candidate stores from the store scoring model, determine the probability information of the user redeeming the rights in the candidate stores; wherein, the enhanced features are obtained by enhancing the representation of the information in the store features of the candidate stores that the user prefers and is suitable for redeeming the rights.

[0234] The sorting module 53 is used to sort the multiple candidate stores according to the probability information of each candidate store.

[0235] In some examples, the store enhancement features are obtained in the following ways:

[0236] Obtain the feature sequence of the historical popular redemption store set corresponding to the benefit, and obtain the feature sequence of the user's historical preferred store set;

[0237] By utilizing the feature sequences of the historical popular redemption store set and the feature sequences of the historical preferred store set, target features that represent user preferences and are suitable for redeeming the benefits are obtained;

[0238] Obtain the store features of the candidate stores, and enhance the representation of the information related to the target features in the store features of the candidate stores to obtain enhanced store features.

[0239] In some examples, the target features include: a first target feature, a second target feature, or a third target feature; wherein,

[0240] The first target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preferred store set;

[0241] The second target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preference store set;

[0242] The third target feature is obtained by fusing the first target feature and the second target feature.

[0243] In some examples, the store rating model includes:

[0244] A first Transformer structure is used to enhance the representation of information related to the feature sequence of the historically popular redemption store set in the feature sequence of the historically preferred store set, and output a first target feature; or...

[0245] A second Transformer structure is used to enhance the representation of information related to the feature sequences of the historical popular redemption stores in the feature sequences of the historical preference store set, and output a second target feature; or...

[0246] The system comprises a first Transformer structure, a second Transformer structure, and a connection layer. The connection layer is connected to the output of the first Transformer structure and to the output of the second Transformer structure. The connection layer is used to concatenate the first target feature output by the first Transformer structure and the second target feature output by the second Transformer structure to output the third target feature.

[0247] In some examples, the feature sequence of the historical popular redemption store set is obtained in the following way:

[0248] Obtain the store characteristics of each store in the historical popular redemption store set;

[0249] The information related to the store characteristics of the candidate stores is enhanced in the store characteristics of each store.

[0250] The enhanced store characteristics of each store are combined into a feature sequence.

[0251] In some examples, the historical popular redemption store set corresponding to the benefit includes: the user's location and the historical popular redemption store set corresponding to the benefit.

[0252] In some examples, the device is also used for:

[0253] The storage module is used to: acquire historical redemption data within a set time range, group and statistically analyze the data according to geographical location and benefit threshold, obtain multiple sets of historical popular redemption stores, acquire the feature sequence of each set of historical popular redemption stores, and store it.

[0254] The scoring module is also used for:

[0255] From the feature sequences of each set of historical popular redemption stores stored in the database, obtain the feature sequence corresponding to the user's location and the benefit threshold.

[0256] In some examples, after obtaining the enhanced features of the candidate stores from the store scoring model, determining the probability information of the user redeeming the benefits in each of the candidate stores includes:

[0257] After obtaining the benefit characteristics of the benefits and the enhanced characteristics of the candidate stores from the store scoring model, the probability information of the user redeeming the benefits in each of the candidate stores is determined;

[0258] The rights and benefits features include basic features that characterize the basic information of the rights and benefits, and / or combined features obtained by combining the basic features with any of the following features: user features of the user, scene features of the current scene, store features of the candidate store, or user behavior features towards the candidate store.

[0259] In some examples, the recall module 51 is also used for:

[0260] Get the user's currently claimed benefits.

[0261] In some examples, the device also includes a transmitting module:

[0262] Extract the set number of candidate stores with the highest probability information from the sorted candidate stores and send them to the client.

[0263] like Figure 6 As shown, Figure 6 This is a block diagram illustrating a store display device according to an exemplary embodiment of this specification, the device comprising:

[0264] The acquisition module 61 is configured to: respond to a store display request, acquire multiple sorted candidate stores sent by the server; the candidate stores are those retrieved by the server based on the user's benefits; after acquiring the enhanced features of the candidate stores from the store scoring model, determine the probability information of the user redeeming the benefits in the candidate stores; and sort the multiple candidate stores according to the probability information of each candidate store; wherein, the enhanced features are obtained by enhancing the representation of the information in the store features of the candidate stores that is preferred by the user and suitable for redeeming the benefits.

[0265] Display module 62 is used to display the multiple sorted candidate stores.

[0266] Optionally, the store enhancement features are obtained in the following way:

[0267] Obtain the feature sequence of the historical popular redemption store set corresponding to the benefit, and obtain the feature sequence of the user's historical preferred store set;

[0268] By utilizing the feature sequences of the historical popular redemption store set and the feature sequences of the historical preferred store set, target features that represent user preferences and are suitable for redeeming the benefits are obtained;

[0269] Obtain the store features of the candidate stores, and enhance the representation of the information related to the target features in the store features of the candidate stores to obtain enhanced store features.

[0270] Optionally, the target feature includes: a first target feature, a second target feature, or a third target feature; wherein,

[0271] The first target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preferred store set;

[0272] The second target feature is obtained by enhancing the representation of information related to the feature sequence of the historical popular redemption store set in the feature sequence of the historical preference store set;

[0273] The third target feature is obtained by fusing the first target feature and the second target feature.

[0274] Optionally, the store rating model includes:

[0275] A first Transformer structure is used to enhance the representation of information related to the feature sequence of the historically popular redemption store set in the feature sequence of the historically preferred store set, and output a first target feature; or...

[0276] A second Transformer structure is used to enhance the representation of information related to the feature sequences of the historical popular redemption stores in the feature sequences of the historical preference store set, and output a second target feature; or...

[0277] The system comprises a first Transformer structure, a second Transformer structure, and a connection layer. The connection layer is connected to the output of the first Transformer structure and to the output of the second Transformer structure. The connection layer is used to concatenate the first target feature output by the first Transformer structure and the second target feature output by the second Transformer structure to output the third target feature.

[0278] Optionally, the feature sequence of the historical popular redemption store set is obtained in the following way:

[0279] Obtain the store characteristics of each store in the historical popular redemption store set;

[0280] The information related to the store characteristics of the candidate stores is enhanced in the store characteristics of each store.

[0281] The enhanced store characteristics of each store are combined into a feature sequence.

[0282] Optionally, the historical popular redemption store set corresponding to the benefit includes: the user's location and the historical popular redemption store set corresponding to the benefit.

[0283] Optionally, obtaining the feature sequence of the historical popular redemption store set corresponding to the rights includes:

[0284] Obtain historical redemption data within a set time range, and group and statistically analyze it according to geographical location and benefit threshold. After obtaining multiple sets of historical popular redemption stores, obtain the feature sequence of each set of historical popular redemption stores and store it.

[0285] From the feature sequences of each set of historical popular redemption stores stored in the database, obtain the feature sequence corresponding to the user's location and the benefit threshold.

[0286] Optionally, after obtaining the enhanced features of the candidate stores from the store scoring model, determining the probability information of the user redeeming the benefits in each of the candidate stores includes:

[0287] After obtaining the benefit characteristics of the benefits and the enhanced characteristics of the candidate stores from the store scoring model, the probability information of the user redeeming the benefits in each of the candidate stores is determined;

[0288] The rights and benefits features include basic features that characterize the basic information of the rights and benefits, and / or combined features obtained by combining the basic features with any of the following features: user features of the user, scene features of the current scene, store features of the candidate store, or user behavior features towards the candidate store.

[0289] Optionally, obtaining user rights includes:

[0290] Get the user's currently claimed benefits.

[0291] Optionally, the sending module is further configured to:

[0292] Extract the set number of candidate stores with the highest probability information from the sorted candidate stores and send them to the client.

[0293] The specific implementation process of the functions and roles of each module in the aforementioned store sorting device / store display method is detailed in the implementation process of the corresponding steps in the aforementioned store sorting method / store display method, and will not be repeated here.

[0294] Accordingly, this specification also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned store sorting method / store display method embodiments.

[0295] Accordingly, embodiments of this specification also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the aforementioned store sorting method / store display method embodiments.

[0296] Accordingly, embodiments of this specification also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned store sorting method / store display method embodiments.

[0297] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution in this specification according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0298] The above embodiments can be applied to one or more computer devices. The computer device is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. The hardware of the computer device includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0299] The computer device can be any electronic product that can interact with the user, such as a personal computer, tablet computer, smartphone, personal digital assistant (PDA), game console, interactive network television (IPTV), smart wearable device, etc.

[0300] The computer equipment may also include network equipment and / or user equipment. The network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers.

[0301] The network in which the computer device is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, and virtual private network (VPN).

[0302] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0303] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0304] The terms "specific example" or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with the embodiments or examples, which are included in at least one embodiment or example of this specification. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0305] Other embodiments of this specification will readily occur to those skilled in the art upon consideration of the specification and practice of the invention claimed herein. This specification is intended to cover any variations, uses, or adaptations that follow the general principles of this specification and include common knowledge or customary techniques in the art not claimed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this specification are indicated by the following claims.

[0306] It should be understood that this specification is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this specification is limited only by the appended claims.

[0307] The above description is merely a preferred embodiment of this specification and is not intended to limit this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of protection of this specification.

Claims

1. A store sorting method, the method comprising: To obtain the user's rights, multiple candidate stores that can be used to redeem the rights are recalled for the user; After obtaining the enhanced features of the candidate stores from the store scoring model, the probability information of the user redeeming the benefits in the candidate stores is determined, and the multiple candidate stores are sorted according to the probability information of each candidate store. The output of the store scoring model includes the product of the predicted click-through rate and the predicted conversion rate, with the product indicating the probability information. At the feature level, the store scoring model includes a basic benefits feature module, an LBS benefits activation sequence module, and a benefits dual-sequence module. The basic rights and benefits feature module is used to obtain basic features that represent the basic information of the rights and benefits, and to combine the basic features with user features, store features, scenario features, and user-store cross features to obtain combined features; The LBS benefits activation sequence module is used to obtain the feature sequence of the historical popular redemption stores corresponding to the user's location and the benefit threshold. Using the store features of the candidate stores as the query and the feature sequence as the key and value, the module uses an attention mechanism to determine the correlation between the store vector of the candidate stores and the store vectors of N historical popular redemption stores in the feature sequence, thereby obtaining the enhanced feature vectors of N historical popular redemption stores, which are then combined into the activated LBS benefits activation sequence. The equity dual-sequence module includes a collaborative attention structure, which comprises a first Transformer structure and a second Transformer structure, wherein... The first Transformer structure uses the user's historical preferred store sequence as the Query and the activated LBS benefit sequence as the Key and Value. It strengthens the representation of information related to the activated LBS benefit sequence in the historical preferred store sequence and outputs the first strengthened feature sequence. The second Transformer structure uses the activated LBS rights activation sequence as the Query and the user's historical preferred store sequence as the Key and Value to enhance the representation of the information related to the historical preferred store sequence in the activated LBS rights activation sequence and output the second enhanced feature sequence. The first enhanced feature sequence and the second enhanced feature sequence are concatenated to obtain a new feature sequence; Using the store features of candidate stores as the Query and the new feature sequence as the Key and Value, the attention mechanism is used to enhance the representation of the information related to the new feature sequence in the store features of candidate stores, thus obtaining the enhanced store features of candidate stores. The store scoring model integrates the features extracted from the basic benefits feature module, the LBS benefits activation sequence module, and the benefits dual sequence module through a feature embedding layer, and then inputs them into multiple fully connected layers to predict click-through rate and conversion rate.

2. The method according to claim 1, further comprising: Obtain historical redemption data within a set time range, and group and statistically analyze it according to geographical location and benefit threshold. After obtaining multiple sets of historical popular redemption stores, obtain the feature sequence of each set of historical popular redemption stores and store it.

3. The method according to claim 1, wherein obtaining the user's rights includes: Get the user's currently claimed benefits.

4. The method according to claim 1, further comprising: Extract the set number of candidate stores with the highest probability information from the sorted candidate stores and send them to the client.

5. A store display method, the method comprising: In response to a store display request, retrieve multiple sorted candidate stores sent by the server; The candidate stores are obtained by the server from the user's rights and interests. For the user, multiple candidate stores that can redeem the rights and interests are recalled. After the store scoring model obtains the store enhancement features of the candidate stores, the probability information of the user redeeming the rights and interests in the candidate stores is determined. Based on the probability information of each candidate store, the multiple candidate stores are sorted. Display the multiple sorted candidate stores; The output of the store scoring model includes the product of the predicted click-through rate and the predicted conversion rate, with the product indicating the probability information. At the feature level, the store scoring model includes a basic benefits feature module, an LBS benefits activation sequence module, and a benefits dual-sequence module. The basic rights and benefits feature module is used to obtain basic features that represent the basic information of the rights and benefits, and to combine the basic features with user features, store features, scenario features, and user-store cross features to obtain combined features; The LBS benefits activation sequence module is used to obtain the feature sequence of the historical popular redemption stores corresponding to the user's location and the benefit threshold. Using the store features of the candidate stores as the query and the feature sequence as the key and value, the module uses an attention mechanism to determine the correlation between the store vector of the candidate stores and the store vectors of N historical popular redemption stores in the feature sequence, thereby obtaining the enhanced feature vectors of N historical popular redemption stores, which are then combined into the activated LBS benefits activation sequence. The equity dual-sequence module includes a collaborative attention structure, which comprises a first Transformer structure and a second Transformer structure, wherein... The first Transformer structure uses the user's historical preferred store sequence as the Query and the activated LBS benefit sequence as the Key and Value. It strengthens the representation of information related to the activated LBS benefit sequence in the historical preferred store sequence and outputs the first strengthened feature sequence. The second Transformer structure uses the activated LBS rights activation sequence as the Query and the user's historical preferred store sequence as the Key and Value to enhance the representation of the information related to the historical preferred store sequence in the activated LBS rights activation sequence and output the second enhanced feature sequence. The first enhanced feature sequence and the second enhanced feature sequence are concatenated to obtain a new feature sequence; Using the store features of candidate stores as the Query and the new feature sequence as the Key and Value, the attention mechanism is used to enhance the representation of the information related to the new feature sequence in the store features of candidate stores, thus obtaining the enhanced store features of candidate stores. The store scoring model integrates the features extracted from the basic benefits feature module, the LBS benefits activation sequence module, and the benefits dual sequence module through a feature embedding layer, and then inputs them into multiple fully connected layers to predict click-through rate and conversion rate.

6. The method according to claim 5, Obtain historical redemption data within a set time range, and group and statistically analyze it according to geographical location and benefit threshold. After obtaining multiple sets of historical popular redemption stores, obtain the feature sequence of each set of historical popular redemption stores and store it.

7. The method according to claim 5, wherein obtaining the user's rights includes: Get the user's currently claimed benefits.

8. The method according to claim 5, further comprising: Extract the set number of candidate stores with the highest probability information from the sorted candidate stores and send them to the client.

9. A store sorting device, the device comprising: The recall module is used to: obtain the user's rights and interests, and recall multiple candidate stores that can redeem the rights and interests for the user; The scoring module is used to: after obtaining the enhanced features of the candidate stores from the store scoring model, determine the probability information of the user redeeming the rights in the candidate stores; The sorting module is used to sort the multiple candidate stores according to the probability information of each candidate store; The output of the store scoring model includes the product of the predicted click-through rate and the predicted conversion rate, with the product indicating the probability information. At the feature level, the store scoring model includes a basic benefits feature module, an LBS benefits activation sequence module, and a benefits dual-sequence module. The basic rights and benefits feature module is used to obtain basic features that represent the basic information of the rights and benefits, and to combine the basic features with user features, store features, scenario features, and user-store cross features to obtain combined features; The LBS benefits activation sequence module is used to obtain the feature sequence of the historical popular redemption stores corresponding to the user's location and the benefit threshold. Using the store features of the candidate stores as the query and the feature sequence as the key and value, the module uses an attention mechanism to determine the correlation between the store vector of the candidate stores and the store vectors of N historical popular redemption stores in the feature sequence, thereby obtaining the enhanced feature vectors of N historical popular redemption stores, which are then combined into the activated LBS benefits activation sequence. The equity dual-sequence module includes a collaborative attention structure, which comprises a first Transformer structure and a second Transformer structure, wherein... The first Transformer structure uses the user's historical preferred store sequence as the Query and the activated LBS benefit sequence as the Key and Value. It strengthens the representation of information related to the activated LBS benefit sequence in the historical preferred store sequence and outputs the first strengthened feature sequence. The second Transformer structure uses the activated LBS rights activation sequence as the Query and the user's historical preferred store sequence as the Key and Value to enhance the representation of the information related to the historical preferred store sequence in the activated LBS rights activation sequence and output the second enhanced feature sequence. The first enhanced feature sequence and the second enhanced feature sequence are concatenated to obtain a new feature sequence; Using the store features of candidate stores as the Query and the new feature sequence as the Key and Value, the attention mechanism is used to enhance the representation of the information related to the new feature sequence in the store features of candidate stores, thus obtaining the enhanced store features of candidate stores. The store scoring model integrates the features extracted from the basic benefits feature module, the LBS benefits activation sequence module, and the benefits dual sequence module through a feature embedding layer, and then inputs them into multiple fully connected layers to predict click-through rate and conversion rate.

10. A store display device, the device comprising: The acquisition module is used to: in response to a store display request, acquire multiple sorted candidate stores sent by the server; The candidate stores are the rights and interests of the user obtained by the server. For the user, multiple candidate stores that can be used to redeem the rights and interests are recalled. After the store scoring model obtains the store enhancement features of the candidate stores, the probability information of the user redeeming the rights and interests in the candidate stores is determined. The multiple candidate stores are sorted according to their probability information. The display module is used to: display the multiple sorted candidate stores; The output of the store scoring model includes the product of the predicted click-through rate and the predicted conversion rate, with the product indicating the probability information. At the feature level, the store scoring model includes a basic benefits feature module, an LBS benefits activation sequence module, and a benefits dual-sequence module. The basic rights and benefits feature module is used to obtain basic features that represent the basic information of the rights and benefits, and to combine the basic features with user features, store features, scenario features, and user-store cross features to obtain combined features; The LBS benefits activation sequence module is used to obtain the feature sequence of the historical popular redemption stores corresponding to the user's location and the benefit threshold. Using the store features of the candidate stores as the query and the feature sequence as the key and value, the module uses an attention mechanism to determine the correlation between the store vector of the candidate stores and the store vectors of N historical popular redemption stores in the feature sequence, thereby obtaining the enhanced feature vectors of N historical popular redemption stores, which are then combined into the activated LBS benefits activation sequence. The equity dual-sequence module includes a collaborative attention structure, which comprises a first Transformer structure and a second Transformer structure, wherein... The first Transformer structure uses the user's historical preferred store sequence as the Query and the activated LBS benefit sequence as the Key and Value. It strengthens the representation of information related to the activated LBS benefit sequence in the historical preferred store sequence and outputs the first strengthened feature sequence. The second Transformer structure uses the activated LBS rights activation sequence as the Query and the user's historical preferred store sequence as the Key and Value to enhance the representation of the information related to the historical preferred store sequence in the activated LBS rights activation sequence and output the second enhanced feature sequence. The first enhanced feature sequence and the second enhanced feature sequence are concatenated to obtain a new feature sequence; Using the store features of candidate stores as the Query and the new feature sequence as the Key and Value, the attention mechanism is used to enhance the representation of the information related to the new feature sequence in the store features of candidate stores, thus obtaining the enhanced store features of candidate stores. The store scoring model integrates the features extracted from the basic benefits feature module, the LBS benefits activation sequence module, and the benefits dual sequence module through a feature embedding layer, and then inputs them into multiple fully connected layers to predict click-through rate and conversion rate.

11. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method of any one of claims 1 to 8.

12. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images