Sorting model training method, sorting method and device based on multi-scene data

Abstract

The embodiments of the present invention provide a sorting model training method, sorting method and device based on multi-scenario data, the training method comprising: constructing sample data; generating sample features based on the scene ID; the feature data comprising single-scenario features and multi-scenario features; the single-scenario features comprising the user's behavior sequence information and behavior statistics in a single scene, and the statistical information of the sorting object in a single scene; the multi-scenario features comprising the user's behavior sequence information and behavior statistics in multiple scenes, and the statistical information of the sorting object in multiple scenes; inputting the sample features into the model, and performing neural network training based on the output labels to obtain a sorting model. The embodiments of the present invention achieve differentiated learning of different scenes and linkage optimization between scenes by performing scene differentiation in the sample and performing model learning through single-scenario features and multi-scenario features, and can solve the problem of sparse target data, thereby effectively improving sorting accuracy.

Description

Technical Field

[0001] Embodiments of the present invention relate to a sorting model training method, a sorting method, and a device based on multi-scene data. Background Art

[0002] Many apps require sorting, most of which is used in search and recommendation scenarios. For example, after searching for properties in a real estate app, a ranking result for the properties is displayed, and the ranking of recommended properties is also provided in the corresponding module. Commodity trading platforms can also provide ranking results for products by searching for a specific product, and the ranking of recommended products is also provided in the corresponding module. Both search and recommendation can be further divided into different scenarios. For example, the homepage of a property app displays a "Recommended for You" module, which shows a list of recommended properties to be sorted. After clicking on a property card to enter the property details, the "Guess You Like" module can be further configured to make further recommendations based on the property being viewed and display the ranking results. Of course, there may be other scenarios in the app that also require sorting functionality. Data from these different scenarios constitute multi-scenario data. Summary of the Invention

[0003] An embodiment of the present invention provides a sorting model training method based on multi-scenario data, comprising: constructing sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID and at least one label; adding feature data on the basis of the scene ID to generate sample features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scenario features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; the sample features are input into a neural network model, and the neural network training is performed with the at least one label as the output label, and the sorting model is obtained after the training is completed.

[0004] According to a method for training a ranking model based on multi-scene data provided by an embodiment of the present invention, the single-scene feature also includes contextual features of a single scene corresponding to the scene ID.

[0005] According to an embodiment of the present invention, a sorting model training method based on multi-scene data is provided, wherein the single-scene features include historical single-scene features and real-time single-scene features; the multi-scene features include historical multi-scene features and real-time multi-scene features; the feature data also includes at least one of user basic features, sorting object basic features, sequence features and cross-features; wherein: the user basic features include the basic attribute features and user portrait features of the user; the sorting object basic features include the preset basic features of the sorting object; the sequence features include the correlation features between the sorting object corresponding to the sorting object ID and the historical behavior sequence information; the cross-features include the cross-statistical features of at least one user basic feature and at least one single-scene feature, the cross-statistical features of at least one user basic feature and at least one multi-scene feature, the cross-statistical features of at least one user basic feature and the sequence features and / or the cross-statistical features of the user ID and the sorting object ID.

[0006] An embodiment of the present invention also provides a sorting method, comprising: obtaining information of a user ID, a scene ID, and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; adding feature data based on the scene ID to generate input features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in preset multiple scenes, and statistical information of the sorting object corresponding to the sorting object ID in the preset multiple scenes; inputting the input features into the sorting model to obtain the score of at least one label of the sorting object in the sorting object list, and sorting the sorting objects according to the score of the at least one label.

[0007] According to a sorting method provided by an embodiment of the present invention, sorting the sorting objects according to the score of the at least one tag includes: obtaining a comprehensive score of the sorting objects according to the score of the at least one tag, and sorting the sorting objects according to the comprehensive score.

[0008] An embodiment of the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements any of the above-described methods for training a sorting model based on multi-scenario data or the steps of the sorting method.

[0009] An embodiment of the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any of the above-mentioned sorting model training methods based on multi-scene data or the steps of the sorting method.

[0010] An embodiment of the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned steps of the sorting model training method based on multi-scene data or the sorting method.

[0011] The embodiments of the present invention provide a sorting model training method, a sorting method, and a device based on multi-scene data. By adding scene IDs to samples to distinguish scenes and performing model learning through single-scene features and multi-scene features, not only can differentiate and learn different scenes, but also can realize linkage optimization between scenes, and can solve the problem of sparse target data, thereby effectively improving sorting accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0013] Figure 1 It is a schematic diagram of modeling technology in some technical solutions;

[0014] Figure 2 It is a schematic diagram of modeling technology in other technical solutions;

[0015] Figure 3 This is one of the flow charts of the method for training a ranking model based on multi-scenario data provided by an embodiment of the present invention;

[0016] Figure 4 It is a schematic diagram of sample data constructed in some technical solutions;

[0017] Figure 5 is a schematic diagram of sample data constructed by the ranking model training method based on multi-scenario data provided by an embodiment of the present invention;

[0018] Figure 6 This is an example of sample data constructed by the ranking model training method based on multi-scenario data provided by an embodiment of the present invention;

[0019] Figure 7 This is the second flow chart of the method for training a ranking model based on multi-scenario data provided by an embodiment of the present invention;

[0020] Figure 8 1 is a flow chart of a sorting method provided by an embodiment of the present invention;

[0021] Figure 9 Schematic diagram of the structure of a sorting model training device based on multi-scenario data provided by an embodiment of the present invention;

[0022] Figure 10 is a schematic structural diagram of a sorting device provided by an embodiment of the present invention;

[0023] Figure 11 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0025] To implement the sorting function for each scenario, it is necessary to build and train a sorting model through a neural network, and then automatically provide the sorting results. The modeling process includes: constructing samples, generating features, building models, and training models. Figure 1 This is a schematic diagram of modeling technology in some technical solutions. Figure 1 As shown in the figure, for a large homepage sample (such as the homepage recommendation scenario), feature engineering is used to extract sample features, which are then aggregated through feature layers (such as the embedding layer). The processed features are then input into a deep model network (such as W&DL) to obtain the trained model output-1. This output-1 can then be used for homepage recommendations. Figure 2 This is a schematic diagram of modeling technology in other technical solutions. Figure 2 For samples with few result pages (e.g., search results with few recommendations), we extract sample features through feature engineering, aggregate them through feature layers (e.g., embedding layers), and then feed the processed features into a deep model network (e.g., DeepFM) to obtain the trained model output-2. This output-2 can then be used to make recommendations for pages with few result pages.

[0026] It can be seen that most of these technical solutions are based on single-scene data for modeling, and construct single-objective or multi-objective models. When users browse within the APP, there is an intrinsic connection between different scenes, and users express different mentalities for different scenes. For example, users are accustomed to using search scenes, and their mentalities for search scenes are relatively strong. Users will be willing to stay and operate more in this scene, but only the information of a single scene is considered during modeling, and each scene is modeled independently, forming an information island, which cannot utilize effective information from other scenes. In addition, there may be a problem of insufficient sample size for some scenes with smaller traffic, which may lead to insufficient training. Most of the multi-objective optimization is based on a single scene for optimization, and single-scene multi-objective optimization will encounter the problem of sparse target data for some targets. It is generally solved by sample expansion and other methods, but it cannot fundamentally solve the problem of sparse target data.

[0027] Existing multi-scenario joint modeling approaches mostly simply overlay and integrate data from various scenarios before creating a unified model. This results in consistent scores for the same item (ranking target) across all scenarios, failing to reflect the user's personalized results for different scenarios. Furthermore, since users may have different behaviors towards the same item in different scenarios, this data introduces noise, impacting model learning.

[0028] Figure 3 This is one of the flow charts of the method for training a ranking model based on multi-scenario data provided by an embodiment of the present invention. Figure 3 As shown, the method includes:

[0029] Step 101: Construct sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID, and at least one tag.

[0030] Some existing multi-scenario joint modeling approaches capture user behavior across multiple scenarios for modeling, but they are unable to effectively differentiate between scenarios or optimize them in a coordinated manner. Existing optimization methods fail to specifically address the differences between scenarios, resulting in a lack of interaction between scenarios and a lack of differentiation in user perception across them.

[0031] The ranking model training method based on multi-scenario data provided by the embodiment of the present invention is to better learn the user's behavior in different scenarios and solve the problems such as information islands caused by single-scenario optimization. By setting single-scenario features and multi-scenario features, and introducing scene IDs to distinguish different scenarios, multiple scenarios are jointly optimized. Through multi-scenario joint modeling, the user's behavior in different scenarios can be learned, and the user's multi-scenario features can be used to perform linkage optimization between scenarios.

[0032] In the process of multi-scene linkage joint modeling, that is, using multi-scene joint training ranking models, it can be multi-scene single-target joint modeling or multi-scene multi-target joint modeling according to needs. In an APP, there are inherent connections between different scenes. For example, various search, recommendation and other scenes are inherently related. These scenes all give recommendation results or search results based on a whole background data. All scenes with inherent connections can be selected for joint modeling. In this case, multi-scene joint modeling can also be called full-scene joint modeling. Of course, it is also possible to select some scenes in the full scene for joint modeling according to needs. The multiple scenes involved in the modeling will achieve linkage optimization.

[0033] Before joint modeling, that is, before using multi-scenario data to train the ranking model in the embodiment of the present invention, sample data is first constructed. Obtain the user's behavioral data in multiple scenarios such as searching and recommending within the App, and construct sample data based on this data. The sample data constructed in the embodiment of the present invention includes a user ID, a ranking object ID, a scene ID, and at least one label. Among them, the user ID is used to distinguish different users, and the ranking object ID is used to distinguish different ranking objects. The scene ID is used to distinguish different scenes. At least one label corresponds to the optimization target, that is, the principle based on which the ranking model is sorted. For example, the ranking object is sorted according to the number of exposures, clicks, business opportunities (im, 400, etc.), attention, sharing, etc., and each principle corresponds to an optimization target. If a ranking principle is set, then a label can be set according to the corresponding optimization target; if multiple ranking principles are set, then multiple labels can be set according to the corresponding multiple optimization targets. Neural networks can achieve joint optimization of single and multiple targets.

[0034] Figure 4 This is a schematic diagram of sample data constructed in some technical solutions. Figure 4 As shown, in some technical solutions, the constructed sample data includes user ID, item ID and Label, where item represents a sorting object.

[0035] Figure 5 Schematic diagram of sample data constructed by the method for training a ranking model based on multi-scenario data provided by an embodiment of the present invention. Figure 5 As shown, the sample data constructed in the embodiment of the present invention includes user ID, item ID, scene ID and Label. Figure 4 The difference between the sample construction methods shown is that the embodiment of the present invention adds a scene ID to the sample data to distinguish the user's behavior data in different scenes. By adding the scene ID, the model can learn the different behaviors of users in multiple scenes.

[0036] The labels in the sample data represent various user behaviors, which are also the multiple objectives that the model optimizes. There are no fixed requirements for the label data format; we'll use the Multi-hot encoding format as an example. For example, if there are four optimization objectives (clicks, leads, followers, and shares), the label length is set to 4, with each position representing a 0 / 1 value, indicating whether a click occurred, a lead occurred, a follow occurred, or a share occurred.

[0037] Figure 6 This is an example of sample data constructed by the method for training a ranking model based on multi-scenario data provided by an embodiment of the present invention. Figure 6 The data format of the sample data is illustrated as follows: Figure 6 As shown, in the first sample, user user_01 clicked on item_01, generated business opportunities, and paid attention to it in scene scene_01, but did not share it. In the second sample, user user_01 only clicked on item_01 in scene scene_02, and did not engage in other behaviors. In the third sample, user user_01 clicked on item_02 in scene scene_02, generated business opportunities, and shared it, but did not pay attention to it. Compared with the first two samples, user user_01 clicked on item_01, but other behaviors were different in different scenes (the first sample was in scene scene_01, and the second sample was in scene scene_02). In some technical solutions, scene information is not added, resulting in noisy data in the samples. The embodiment of the present invention can distinguish samples and learn user behaviors in different scenes by adding scene IDs.

[0038] In the sample construction process for multi-scenario modeling, this embodiment of the present invention adds scenario ID information (for different recommendation scenarios, search scenarios, etc.). Users may label the same ranking object differently in different scenarios. By adding scenario ID information to differentiate samples, the model can distinguish samples from different scenarios and learn behaviors in different scenarios.

[0039] Step 102: Add feature data based on the scene ID to generate sample features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include the behavior sequence information and behavior statistical information of the user corresponding to the user ID in the single scene corresponding to the scene ID, and the statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include the behavior sequence information and behavior statistical information of the user corresponding to the user ID in multiple preset scenes, and the statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes.

[0040] Based on the scene ID, feature data is added to generate sample features, which serve as model input during training. Sample features include the added feature data and scene ID information. Once the ranking model is built, it can be applied to different scenarios. Therefore, the scene ID is required as model input to learn data from different scenarios during model training.

[0041] During the feature engineering phase (i.e., generating sample features), for each sample data point, not only are features for the ranking object constructed for multiple pre-defined scenarios (multi-scenario features), but also features for the ranking object in the current scenario (single-scenario features). This results in different model prediction scores for the same user and the same ranking object in different scenarios, reflecting the varying user awareness and engagement with different scenarios.

[0042] Single-scenario features include user behavior sequence information (e.g., the sequence of clicks on sorted objects) and behavioral statistics (e.g., number of clicks, conversion rate, etc.) in the current scenario. They can also include statistical information about the sorted objects in the current scenario, such as exposure, clicks, number of leads, conversion rate, etc. Data from different scenarios can be distinguished by the scenario ID and unique features (single-scenario features) in the sample features. By using single-scenario features, we can further learn about user behavior in each scenario.

[0043] Multi-scenario features include user behavior sequence information and behavior statistics across multiple pre-defined scenarios. They can also include statistical information for sorted objects across multiple pre-defined scenarios, such as impressions, clicks, number of leads, and conversion rates. Pre-defined scenarios include scenarios corresponding to scenario IDs and other intrinsically related scenarios. Using multi-scenario features, you can capture user behavior in other scenarios and apply it to the current scenario, enabling cross-scenario learning and optimization.

[0044] By adding single-scene features and full-scene features, we can learn about users' mindsets in different scenarios. The same user and the same sorting object can get different results in different scenarios. At the same time, the application of multi-scene features can enable each scenario to interact with each other and perform linkage optimization between scenarios.

[0045] Step 103: Input the sample features into a neural network model, perform neural network training using the at least one label as an output label, and obtain the ranking model after the training is completed.

[0046] To facilitate model learning, for scene-distinguishing features, if a feature only appears in one scene, the feature can be left blank in other scenes. The generated sample features are input into the neural network model, and the neural network training is performed with at least one label in the sample data as the output label. After the training end conditions are met, a trained sorting model is obtained. The obtained sorting model is used to sort the sorting objects according to the optimization goals (corresponding to the set labels). If it is multi-objective optimization, multiple labels are required. Taking the optimization of the four goals of clicks, business opportunities, attention, and sharing as an example, the model output part needs to output four scores, which respectively represent the probability values ​​of the four goals.

[0047] The neural network model used in training can be directly integrated into existing mature model structures, such as W&DL, DeepFM, and DNN. Existing mature optimization algorithms, such as SGD and Adam, can also be directly applied during the training process.

[0048] In the embodiment of the present invention, scene IDs and distinguishing single-scene features are used, which can enable the same user to obtain different behavior prediction scores for the same ranking object in different scenarios. Moreover, by adding multi-scene features, the data of other scenes are applied to the current scene to achieve linkage optimization between scenes. In addition, the present invention can combine multi-objective optimization with multi-scene optimization. Based on the advantages of multi-scene modeling, the user's behavior in multiple scenes is integrated to learn and optimize multiple goals as a whole, avoiding the problem of data sparsity in single-scene multi-objective optimization. The overall modeling outputs the scores of the ranking objects corresponding to different goals of the current user, which are then used to sort the ranking objects.

[0049] The sorting model training method based on multi-scene data provided by the embodiment of the present invention distinguishes scenes by adding scene IDs in samples, and learns the model through single-scene features and multi-scene features. It can not only realize the differentiated learning of different scenes, but also realize the linkage optimization between scenes, and can solve the problem of sparse target data.

[0050] According to a method for training a ranking model based on multi-scene data provided by an embodiment of the present invention, the single-scene feature also includes contextual features of a single scene corresponding to the scene ID.

[0051] For example, contextual features for a single scenario can be attribute features of the sorting objects in the current scenario. For example, in the "Guess What You'd Like" recommendation scenario on the property details page, further recommendations are made based on the properties displayed on the current property details page (i.e., the sorting objects in the current scenario). Contextual features can also be search terms and search conditions in search scenarios. These search terms and search conditions further limit the scope of the sorting objects to be sorted.

[0052] Not all single-scene features in all scenarios contain contextual features. For example, the "Recommendations for You" section on the homepage of a real estate app may not have contextual features. For scenarios without contextual features, the corresponding part of the single-scene feature can be left blank.

[0053] The multi-scene data-based sorting model training method provided by the embodiment of the present invention enriches sample features by further including the context features of the current scene in the single-scene features, which is conducive to improving the accuracy of model sorting.

[0054] According to an embodiment of the present invention, a sorting model training method based on multi-scene data is provided, wherein the single-scene features include historical single-scene features and real-time single-scene features; the multi-scene features include historical multi-scene features and real-time multi-scene features; the feature data also includes at least one of user basic features, sorting object basic features, sequence features and cross-features; wherein: the user basic features include the basic attribute features and user portrait features of the user; the sorting object basic features include the preset basic features of the sorting object; the sequence features include the correlation features between the sorting object corresponding to the sorting object ID and the historical behavior sequence information; the cross-features include the cross-statistical features of at least one user basic feature and at least one single-scene feature, the cross-statistical features of at least one user basic feature and at least one multi-scene feature, the cross-statistical features of at least one user basic feature and the sequence features and / or the cross-statistical features of the user ID and the sorting object ID.

[0055] To improve the accuracy of model ranking, feature data can be further enriched for each sample data. For example, single-scene features include historical single-scene features and real-time single-scene features, and multi-scene features include historical multi-scene features and real-time multi-scene features. Feature data can further include at least one of user basic features, sorting object basic features, sequence features, and cross-features. The following uses the real estate app as an example to introduce the above feature data:

[0056] The historical single-scene feature includes the historical behavior sequence information and historical behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, as well as the historical statistics of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID. The real-time single-scene feature includes the real-time behavior sequence information and real-time behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, as well as the real-time statistics of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID. The historical multi-scene feature includes the historical behavior sequence information and historical behavior statistics of the user corresponding to the user ID in multiple scenes corresponding to the scene ID, as well as the historical statistics of the sorting object corresponding to the sorting object ID in multiple scenes corresponding to the scene ID. The real-time multi-scene feature includes the real-time behavior sequence information and real-time behavior statistics of the user corresponding to the user ID in multiple scenes corresponding to the scene ID, as well as the real-time statistics of the sorting object corresponding to the sorting object ID in multiple scenes corresponding to the scene ID. Since user behavior is not only related to historical behavior but also affected by current behavior, considering both historical data and real-time data (such as the user's real-time click count, business opportunities, browsing time, etc., and the real-time exposure count, click count, business opportunities, etc. of the sorting objects) when constructing single-scene features and multi-scene features is conducive to improving the accuracy of model predictions.

[0057] Basic user characteristics include basic user attributes and user profiles. Basic user attributes include basic information such as work city, whether a user is a new user, number of app uses, and activity level. For example, for a real estate app, user profiles might include user preferences for price, area, and room type.

[0058] The basic features of the sorting object include preset basic features of the sorting object, such as the location information, price, area, rooms and other attribute information of the property, and statistical information such as the number of exposures, number of clicks, popularity, conversion rate, etc.

[0059] Sequence features include the correlation between the sorting object corresponding to the sorting object ID and historical behavior sequence information. For example, obtain the user's behavior sequence for the past 10 days and check whether the sorting object in the current sample appears in the user's behavior sequence. If so, the mark is set to 1; if not, the mark is set to 0.

[0060] Cross-features are cross-statistical features of at least two features, which are used to explicitly represent the association between different features, thereby helping neural network learning to improve the accuracy of model ranking. Cross-features include, for example, cross-statistical features of at least one basic user feature and at least one single-scene feature, such as the cross-statistical features of certain basic user features of the user and the user's behavior sequence in the current scene, and the cross-statistical features of certain basic user features of the user and the statistical information of the ranking object in the current scene. For example, for users with male gender and age between 30 and 35, the cross-features of different statistical feature thresholds (exposure, click) for the same ranking object. This will make the model predict different results for different groups of people and the same ranking object, and perform linkage optimization between scenes.

[0061] Cross-features can also include cross-statistical features of at least one user basic feature and at least one multi-scenario feature, such as cross-statistical features of certain user basic features of a user and the user's behavior sequence in multiple preset scenarios, and cross-statistical features of certain user basic features of a user and statistical information of sorting objects in multiple preset scenarios.

[0062] The cross-feature may also include a cross-statistical feature of at least one basic feature of the user and a sequence feature, that is, a sequence feature when the user has a certain basic feature.

[0063] The cross-features may also include cross-statistical features of the user ID and the sorting object ID. The cross-statistical features of the user ID and the sorting object ID may indicate whether the user corresponding to the user ID and the sorting object corresponding to the sorting object ID have ever been associated, such as whether a certain user has viewed a certain property.

[0064] The obtained feature data can be processed as needed, such as outlier processing, normalization, and one-hot encoding. For example, outlier processing and normalization can be performed on continuous features (such as price, area, and number of times), while one-hot encoding can be performed on discrete features (such as urban areas and residential areas).

[0065] The following is the sample data in the real estate APP.<user_01,item_01,scene_01> Taking 1,1,1,0 as an example (in this embodiment, item represents a property), the feature data provided by the embodiments of the present invention is further explained. In real estate applications, multi-scenario generally refers to each recommendation that a user can browse within the app (such as the "Recommended for You" module on the app homepage, the "Guess You Like" module on the property / property details page, etc.), search (searching for neighborhoods / properties, finding properties on a map, etc.), and other matching scenarios. Multi-targets include but are not limited to user clicks, business opportunities, following, sharing, etc. within the app.

[0066] In this example, the user ID is user_01, the sorting object ID is item_01, the scene ID is scene_01, and the labels 1, 1, 1, 0 indicate that a click occurred, a business opportunity occurred, a follow occurred, and no sharing occurred. These labels are used for training and are not used to generate feature data. Based on this sample data, the generated feature data is shown in Table 1.

[0067] Table 1 Example of characteristic data

[0068]

[0069] by Figure 6 The first two samples are used as an example to illustrate that after adding the above features,<user_01,item_01> The single-scene features of different scenes are quite different, which can ensure that different information is learned in scene_01 and scene_02. At the same time, for the same user, the same multi-scene features can ensure interaction between multiple scenes and perform joint optimization.

[0070] The sorting model training method based on multi-scenario data provided by an embodiment of the present invention is further beneficial to improving the sorting accuracy of the sorting model by including historical single-scenario features and real-time single-scenario features in single-scenario features, and including historical multi-scenario features and real-time multi-scenario features in multi-scenario features, and further including at least one of user basic features, sorting object basic features, sequence features and cross-features in the feature data.

[0071] Figure 7 This is the second flow chart of the method for training a ranking model based on multi-scenario data provided by an embodiment of the present invention. Figure 7 As shown, the training of the ranking model includes sample construction, feature construction and processing, model training and model output. In the sample construction stage, sample data is generated according to different scenarios, among which "large homepage sample", "few result page sample", "no result page sample", "details page sample" and "search scenario sample" all represent samples of different scenarios. "Few result page sample" can correspond to the scenario of recommendation when there are relatively few search results. "No result sample" can correspond to the scenario of recommendation when there are no search results. Sample merging means that sample data can be merged into data tables, etc. The purpose of sample deduplication is to improve the reliability of the sample. For example, if there are two clicks on the same ranking object at the same time, or continuous clicks on the same ranking object in a very short time, the number of clicks is taken as 1. "Deconfliction" means that if the same house is exposed (displayed on the page) and clicked, it is considered that a click has occurred.

[0072] During the feature construction phase, feature data including full-scene features, individual scene features, sequence features, and cross-features are constructed, forming a sample feature input model at the shared feature layer. During the model training phase, the model can be trained using existing neural networks such as W&DL, DeepFM, and DNN. After training, the model is output for sorting applications.

[0073] In order to jointly optimize multiple scenarios, the embodiments of the present invention construct single-scenario features and multi-scenario features. Single-scenario features are features of users and sorting objects only in a single scenario, which are used to learn user behaviors in different scenarios, while multi-scenario features (which can be full-scenario features) are features of users and sorting objects in multiple preset scenarios. The addition of multi-scenario features can enable the model to learn the linkage of multiple scenarios and perform joint optimization.

[0074] Figure 8 : is a flow chart of the sorting method provided by an embodiment of the present invention. The sorting method can use the sorting model obtained by the above training to sort the sorting objects. Figure 8 As shown, the method includes:

[0075] Step S1: Obtain information of a user ID, a scene ID, and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects.

[0076] First, obtain the user ID, scene ID, and the sorting object IDs of the multiple sorting objects to be displayed. This means obtaining the user ID, scene ID, and sorting object ID information. This information has the same meaning as the user ID, scene ID, and sorting object ID in the sample data during training.

[0077] The sorting module can sort the sorting objects in the sorting object list. In the application, the number of sorting objects to be sorted is huge. In order to improve the efficiency of online processing, a part of the sorting objects can be recalled offline first, and then the recalled sorting objects can be sorted online. In this way, in the offline recall stage, the sorting objects to be displayed are the full set of sorting objects saved in the APP background. In the online sorting stage, the sorting objects to be displayed are some of the recalled sorting objects. The recall stage also screens the sorting objects by sorting. The sorting model trained by the embodiment of the present invention can be applied to the offline recall and online sorting stages. Of course, the recall stage can also use other methods for recall, such as recall based on multi-way recall.

[0078] Step S2: Add feature data based on the scene ID to generate input features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include the behavior sequence information and behavior statistical information of the user corresponding to the user ID in the single scene corresponding to the scene ID, and the statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include the behavior sequence information and behavior statistical information of the user corresponding to the user ID in multiple preset scenes, and the statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes.

[0079] Feature data is added based on the scene ID to generate input features. The method of generating input features is the same as the method of generating sample features. That is, the feature data includes single-scene features and multi-scene features; single-scene features include the behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and the statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; multi-scene features include the behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and the statistical information of the sorting object corresponding to the sorting object ID in multiple preset scenes.

[0080] Step S3: input the input feature into a ranking model to obtain a score of at least one label of the ranking object in the ranking object list, and sort the ranking objects according to the score of the at least one label.

[0081] Input the input features to the sorting model, obtain the score of at least one label of the sorting objects in the sorting object list according to the label situation during training, and sort the sorting objects according to the score of at least one label.

[0082] Taking real estate applications as an example, in the recall phase, offline batch prediction is performed to determine each user's scores for all properties in each scenario. After sorting by scores, candidate properties for each user's different goals in each scenario are obtained. Properties with high scores are those where the model predicts the user is most likely to engage in the target behavior.

[0083] In the sorting stage: After the user recalls a series of listings in different scenarios, a list of listings to be displayed is created based on the recalled listings. Feature data is added based on the scenario ID to generate input features. The input features are input into the model to obtain the score of each listing. The listings are then sorted according to their scores as the sorting result for the current scenario.

[0084] For example, a user may enter different scenarios (Search 1, Recommendation 1, Recommendation 2). Let's assume the user enters Recommendation 1. At this point, we have obtained the user's basic information, the scenario they entered, and the information about the properties to be displayed. Based on these inputs, we construct the feature information required by the model. After inputting this feature information into the model, we calculate the score for each property to be displayed. Each property is displayed in Recommendation 1 based on its score. If the user enters Recommendation 2, the above steps will be repeated, but the scores for the same properties will be different. This is determined by the sample construction and feature construction.

[0085] The present invention can realize the joint modeling of multiple scenarios and multiple objectives. When modeling multiple scenarios, the user's behavior in a single scenario is learned by adding scenario ID information to the sample and using the single-scene features of each scenario. At the same time, the multi-scene features are used to optimize the current scene in combination with other scene information to achieve the linkage optimization effect between scenes. Based on the behavioral sequence information and behavioral statistical information of multiple scenarios and the joint modeling of multiple scenarios, multi-objective optimization is applied in the output stage, and multi-objective optimization is performed from the perspective of multiple scenarios, avoiding problems such as data sparsity when performing multi-objective optimization on a single scenario, and ultimately improving the overall indicators of multiple scenarios. At the same time, the same user can score the same sorting object differently in different scenarios, reflecting the different mindsets of users for different scenarios and the personalized characteristics of users for different scenarios.

[0086] The sorting method provided by the embodiment of the present invention distinguishes scenes by adding scene IDs to input features, and sorts the sorting objects to be displayed by learning single-scene features and multi-scene features. It can not only realize the differentiated learning of different scenes, but also realize the linkage optimization between scenes, and solve the problem of sparse target data, thereby effectively improving the sorting accuracy.

[0087] A sorting method provided according to an embodiment of the present invention is characterized in that sorting the sorting objects according to the scores of the at least one tag includes: obtaining a comprehensive score of the sorting object according to the scores of the at least one tag, and sorting the sorting objects according to the comprehensive score.

[0088] When training a ranking model with multiple labels, ranking using the ranking model can yield scores for multiple labels, each corresponding to an optimization objective. A comprehensive score for the ranking object can be obtained by weighted summing the scores of at least one label or by using an exponential power algorithm (e.g., taking the scores of each label to an exponential power and then multiplying them, or taking the scores of each label to an exponential power and then adding them together, where the specific exponential power can be set based on the label), and the ranking objects can be sorted based on the comprehensive score. The weights used in the weighted summation can be set based on the importance of the optimization objective.

[0089] Based on the optimization goal of the current scenario (determined by the tag settings to determine whether the optimization goal is a single goal or multiple goals), the corresponding output scores are taken and sorted from high to low according to the scores as the sorting result of the current scenario. For example, if some optimization goals are not of concern, the weight can be set to 0.

[0090] The ranking method provided by the embodiment of the present invention obtains a comprehensive score of the ranking object based on the score of at least one tag, and ranks the ranking objects according to the comprehensive score, thereby improving the flexibility and accuracy of ranking during multi-objective optimization.

[0091] The following describes a sorting model training device based on multi-scenario data provided by an embodiment of the present invention. The sorting model training device based on multi-scenario data described below and the sorting model training method based on multi-scenario data described above can refer to each other.

[0092] Figure 9 Schematic diagram of the structure of the sorting model training device based on multi-scene data provided by an embodiment of the present invention. Figure 9As shown, the sorting model training device includes a sample data construction module 10, a sample feature generation module 20 and a training module 30, wherein: the sample data construction module 10 is used to: construct sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID and at least one label; the sample feature generation module 20 is used to: add feature data on the basis of the scene ID to generate sample features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include the behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and the statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include the behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and the statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; the training module 30 is used to: input the sample features into the neural network model, perform neural network training with the at least one label as the output label, and obtain the sorting model after the training is completed.

[0093] The sorting model training device based on multi-scene data provided by an embodiment of the present invention distinguishes scenes by adding scene IDs in samples, and performs model learning through single-scene features and multi-scene features. It can not only realize the differentiated learning of different scenes, but also realize the linkage optimization between scenes, and can solve the problem of sparse target data.

[0094] According to an embodiment of the present invention, a sorting model training device based on multi-scene data is provided, wherein the single-scene feature further includes a context feature of the single scene corresponding to the scene ID.

[0095] The multi-scene data-based sorting model training device provided by the embodiment of the present invention enriches sample features by further including the context features of the current scene in the single-scene features, which is conducive to improving the accuracy of model sorting.

[0096] According to an embodiment of the present invention, a sorting model training device based on multi-scene data is provided, wherein the single-scene features include historical single-scene features and real-time single-scene features; the multi-scene features include historical multi-scene features and real-time multi-scene features; the feature data also includes at least one of user basic features, sorting object basic features, sequence features and cross-features; wherein: the user basic features include the basic attribute features and user portrait features of the user; the sorting object basic features include the preset basic features of the sorting object; the sequence features include the correlation features between the sorting object corresponding to the sorting object ID and the historical behavior sequence information; the cross-features include the cross-statistical features of at least one user basic feature and at least one single-scene feature, the cross-statistical features of at least one user basic feature and at least one multi-scene feature, the cross-statistical features of at least one user basic feature and the sequence features and / or the cross-statistical features of the user ID and the sorting object ID.

[0097] The sorting model training device based on multi-scenario data provided by an embodiment of the present invention is further beneficial to improving the sorting accuracy of the sorting model by including historical single-scenario features and real-time single-scenario features in single-scenario features, and including historical multi-scenario features and real-time multi-scenario features in multi-scenario features, and further including at least one of user basic features, sorting object basic features, sequence features and cross-features in the feature data.

[0098] Figure 10 Schematic diagram of the structure of the sorting device provided by the embodiment of the present invention. Figure 10 As shown, the sorting device includes an information acquisition module 100, an input feature generation module 200 and a sorting module 300, wherein: the information acquisition module 100 is used to: obtain information of a user ID, a scene ID and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; the input feature generation module 200 is used to: add feature data on the basis of the scene ID to generate input features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in a single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in a single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; the sorting module 300 is used to: input the input features into the sorting model, obtain the score of at least one label of the sorting object in the sorting object list, and sort the sorting objects according to the score of the at least one label.

[0099] The sorting device provided by the embodiment of the present invention distinguishes scenes by adding scene IDs to input features, and sorts the sorting objects to be displayed by learning single-scene features and multi-scene features. It can not only realize the differentiated learning of different scenes, but also realize the linkage optimization between scenes, and can solve the problem of sparse target data, thereby effectively improving the sorting accuracy.

[0100] According to a sorting device provided by an embodiment of the present invention, the sorting module 300, when used to sort the sorting objects according to the score of the at least one tag, is specifically used to: obtain a comprehensive score of the sorting object according to the score of the at least one tag, and sort the sorting objects according to the comprehensive score.

[0101] The ranking device provided by the embodiment of the present invention obtains a comprehensive score of the ranking object according to the score of at least one tag, and ranks the ranking objects according to the comprehensive score, thereby improving the flexibility and accuracy of ranking during multi-objective optimization.

[0102] Figure 11 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 11As shown, the electronic device may include: a processor (processor) 1110, a communication interface (Communications Interface) 1120, a memory (memory) 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140. The processor 1110 can call the logic instructions in the memory 1130 to execute a sorting model training method based on multi-scene data or execute a sorting method, which includes: constructing sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID and at least one label; adding feature data based on the scene ID to generate sample features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include the behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and the statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include the behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and the statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; the sample features are input into the neural network model, and the neural network training is performed with the at least one label as the output label, and the sorting model is obtained after the training is completed. The sorting method includes: obtaining information of a user ID, a scene ID and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; adding feature data on the basis of the scene ID to generate input features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; inputting the input features into a sorting model to obtain a score of at least one label of the sorting object in the sorting object list, and sorting the sorting objects according to the score of the at least one label.

[0103] In addition, the logic instructions in the above-mentioned memory 1130 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0104] On the other hand, an embodiment of the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a sorting model training method based on multi-scenario data or execute a sorting method. The sorting model training method includes: constructing sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID and at least one label; adding feature data on the basis of the scene ID to generate sample features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scenario features include behavior sequence information and behavior statistics of the user corresponding to the user ID in preset multiple scenes, and statistical information of the sorting object corresponding to the sorting object ID in the preset multiple scenes; the sample features are input into a neural network model, and the neural network training is performed with the at least one label as the output label. The sorting model is obtained after the training is completed. The sorting method includes: obtaining information of a user ID, a scene ID and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; adding feature data on the basis of the scene ID to generate input features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; inputting the input features into a sorting model to obtain a score of at least one label of the sorting object in the sorting object list, and sorting the sorting objects according to the score of the at least one label.

[0105] On the other hand, an embodiment of the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a sorting model training method based on multi-scenario data or a sorting method, the sorting model training method comprising: constructing sample data; wherein the sample data comprises a user ID, a sorting object ID, a scene ID and at least one label; adding feature data based on the scene ID to generate sample features; wherein the feature data comprises single-scene features and multi-scene features; the single-scene features comprise behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, as well as statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scenario features comprise behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, as well as statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; the sample features are input into a neural network model, and the neural network training is performed with the at least one label as the output label, and the sorting model is obtained after the training is completed. The sorting method includes: obtaining information of a user ID, a scene ID and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; adding feature data on the basis of the scene ID to generate input features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; inputting the input features into a sorting model to obtain a score of at least one label of the sorting object in the sorting object list, and sorting the sorting objects according to the score of the at least one label.

[0106] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0107] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A ranking model training method based on multi-scene data, characterized in that: include: Constructing sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID and at least one tag; Feature data is added on the basis of the scene ID to generate sample features; wherein the feature data includes single-scene features and multi-scene features; the single-scene features only include features corresponding to the scene ID, and the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; wherein the preset multiple scenes include the scene corresponding to the scene ID and other scenes that are intrinsically related to the scene corresponding to the scene ID; The sample features are input into a neural network model, and the neural network is trained using the at least one label as an output label, and the ranking model is obtained after the training is completed.

2. The method for training a ranking model based on multi-scenario data according to claim 1, characterized in that: The single scene feature also includes context features of a single scene corresponding to the scene ID.

3. The method for training a ranking model based on multi-scenario data according to claim 1, characterized in that: The single-scene feature includes a historical single-scene feature and a real-time single-scene feature; the multi-scene feature includes a historical multi-scene feature and a real-time multi-scene feature; the feature data also includes at least one of a user basic feature, a sorting object basic feature, a sequence feature, and a cross feature; wherein: The user basic characteristics include the user's basic attribute characteristics and user portrait characteristics; The sorting object basic features include preset basic features of the sorting object; The sequence feature includes a correlation feature between the sorting object corresponding to the sorting object ID and the historical behavior sequence information; The cross-features include cross-statistical features of at least one user basic feature and at least one single-scene feature, cross-statistical features of at least one user basic feature and at least one multi-scene feature, cross-statistical features of at least one user basic feature and the sequence feature, and / or cross-statistical features of the user ID and the sorting object ID.

4. A ranking method based on the training method according to any one of claims 1 to 3, characterized in that: include: Obtaining information of a user ID, a scene ID, and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; Feature data is added on the basis of the scene ID to generate input features; wherein, the feature data includes single-scene features and multi-scene features; the single-scene features only include features corresponding to the scene ID, and the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; wherein, the preset multiple scenes include the scene corresponding to the scene ID and other scenes that are intrinsically related to the scene corresponding to the scene ID; The input feature is input into a ranking model to obtain a score of at least one label of the ranking object in the ranking object list, and the ranking objects are ranked according to the score of the at least one label.

5. The sorting method according to claim 4, characterized in that: Sorting the sorting objects according to the score of the at least one tag includes: A comprehensive score of the sorting object is obtained according to the score of the at least one tag, and the sorting object is sorted according to the comprehensive score.

6. A sorting model training device based on multi-scene data, characterized in that: include: A sample data construction module is used to: construct sample data; wherein the sample data includes a user ID, a sorting object ID, a scene ID and at least one tag; A sample feature generation module is used to: add feature data on the basis of the scene ID to generate sample features; wherein, the feature data includes single-scene features and multi-scene features; the single-scene features only include features corresponding to the scene ID, and the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; wherein, the preset multiple scenes include the scene corresponding to the scene ID and other scenes that are intrinsically related to the scene corresponding to the scene ID; a training module is used to: input the sample features into a neural network model, perform neural network training with the at least one label as the output label, and obtain the sorting model after the training is completed.

7. A sorting device based on the training device according to claim 6, characterized in that: include: An information acquisition module is used to: acquire information of a user ID, a scene ID, and a sorting object list to be displayed; wherein the sorting object list includes sorting object IDs of multiple sorting objects; An input feature generation module is used to: add feature data on the basis of the scene ID to generate input features; wherein, the feature data includes single-scene features and multi-scene features; the single-scene features only include features corresponding to the scene ID, and the single-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in the single scene corresponding to the scene ID, and statistical information of the sorting object corresponding to the sorting object ID in the single scene corresponding to the scene ID; the multi-scene features include behavior sequence information and behavior statistics of the user corresponding to the user ID in multiple preset scenes, and statistical information of the sorting object corresponding to the sorting object ID in the multiple preset scenes; wherein, the multiple preset scenes include the scene corresponding to the scene ID and other scenes that are intrinsically related to the scene corresponding to the scene ID; The sorting module is used to: input the input feature into the sorting model, obtain the score of at least one label of the sorting object in the sorting object list, and sort the sorting objects according to the score of the at least one label.

8. An electronic 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 program, the steps of the sorting model training method based on multi-scene data as described in any one of claims 1 to 3 or the steps of the sorting method as described in any one of claims 4 to 5 are implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the sorting model training method based on multi-scene data as described in any one of claims 1 to 3 or the steps of the sorting method as described in any one of claims 4 to 5 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the sorting model training method based on multi-scene data as described in any one of claims 1 to 3 or the steps of the sorting method as described in any one of claims 4 to 5 are implemented.

Images