Object ordering method, device, and storage medium

Abstract

This application discloses an object sorting method, device, and storage medium, belonging to the field of computer technology. The method includes: acquiring interaction behavior data of a first object in multiple business scenarios; determining multi-dimensional features of the first object based on the interaction behavior data; fusing the multi-dimensional features with the original associated features of the first object to obtain fused features; inputting the fused features into a multi-objective sorting model to obtain a sorting score for the first object, and sorting the first object according to the sorting score. This solution effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the sorting results.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an object sorting method, device and storage medium. Background Technology

[0002] In object search scenarios, users search for objects by entering search terms. Ranking algorithms rank search results based on factors such as the relevance of candidate objects to the search terms and the object's own popularity to improve user experience. Taking audio objects as an example, accurate and efficient ranking results are a core element for music platforms to meet user needs and improve retention rates.

[0003] Currently, existing music search ranking technologies typically build ranking models based on features such as the matching degree between search terms and candidate songs, and historical click rates. For example, by statistically analyzing the textual matching degree between search terms and song titles and artists, and combining global indicators such as the total number of plays and favorites of the songs, machine learning models are used to score and rank candidate songs.

[0004] However, the above solutions rely solely on the matching features between search terms and songs, as well as global popularity metrics, making it difficult to fully capture the true performance of songs in different business scenarios. When the search terms themselves lack sufficient information or the candidate songs are new or unpopular songs, the feature sparsity problem becomes particularly prominent, leading to biased ranking results. Summary of the Invention

[0005] This application provides an object sorting method, device, and storage medium, effectively solving the problems of limited features, data bias, and poor timeliness in existing technologies, and improving the accuracy of sorting results. The technical solution is as follows: According to one aspect of this application, an object sorting method is provided, the method comprising: Acquire interactive behavior data of a first object in multiple business scenarios, including search-related scenarios and non-search-related scenarios; Based on the interaction behavior data, the multi-dimensional features of the first object are determined. The multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data. The multi-dimensional features are fused with the original association features of the first object to obtain fused features, wherein the original association features are used to indicate the association relationship between the first object and at least one search term; The fused features are input into a multi-objective ranking model to obtain a ranking score for the first object, and the first object is ranked according to the ranking score.

[0006] According to another aspect of this application, an object sorting apparatus is provided, the apparatus comprising: The acquisition module is used to acquire interactive behavior data of the first object in multiple business scenarios, including search-related scenarios and non-search-related scenarios. The determination module is used to determine the multi-dimensional features of the first object based on the interaction behavior data. The multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data. The fusion module is used to fuse the multi-dimensional features with the original association features of the first object to obtain fused features, wherein the original association features are used to indicate the association relationship between the first object and at least one search term; The sorting module is used to input the fused features into a multi-objective sorting model to obtain a sorting score for the first object, and to sort the first object according to the sorting score.

[0007] According to another aspect of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the object sorting method as described above.

[0008] According to another aspect of this application, a computer-readable storage medium is provided, wherein a computer program is stored therein, the computer program being loaded and executed by a processor to implement the object sorting method as described above.

[0009] According to another aspect of this application, a computer program product is provided, comprising a computer program executed by a processor to implement the object sorting method provided in various alternative implementations of the above aspects.

[0010] This application provides an object ranking scheme. By acquiring interaction behavior data of a first object in multiple business scenarios, it determines multi-dimensional features including numerical features, proportional features, and multi-temporal granular features. These features are then fused with the original association features and input into a multi-objective ranking model. The multi-temporal granular features simultaneously reflect the medium-to-long-term trend and real-time changes in object popularity, enhancing the timeliness of the ranking. The multi-objective ranking model comprehensively optimizes multiple objectives, making the ranking results more consistent with users' actual needs, thus improving the overall quality and stability of the ranking effect. This scheme effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the ranking results. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the implementation environment of an object sorting method provided in an embodiment of this application; Figure 2 This is a flowchart of an object sorting method provided according to an embodiment of this application; Figure 3 This is a flowchart illustrating another object sorting method provided according to an embodiment of this application; Figure 4 This is a flowchart illustrating another object sorting method provided according to an embodiment of this application; Figure 5 This is a flowchart illustrating another object sorting method provided according to an embodiment of this application; Figure 6 This is a flowchart of a multi-objective ranking model provided according to an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an object sorting device according to an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device provided according to an embodiment of this application.

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0015] To enable those skilled in the art to better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0016] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0017] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the media resources involved in this application were obtained with full authorization.

[0018] Figure 1 This is a schematic diagram illustrating the implementation environment of an object sorting method provided in an embodiment of this application. See also... Figure 1 The implementation environment specifically includes: terminal device 101 and server 102. Terminal device 101 can be connected to server 102 via wireless network or wired network.

[0019] Terminal device 101 can be at least one of the following: smartphone, smartwatch, desktop computer, laptop, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), and laptop computer. An application can be installed and run on terminal device 101. This application is compatible with various media content platforms, covering scenarios such as music playback, online singing, live streaming, and on-demand video streaming, and can be in the form of a standalone app or mini-program. This application is associated with server 102, which provides background services to terminal device 101.

[0020] Terminal device 101 can refer to one of a plurality of terminal devices. This embodiment uses terminal device 101 as an example. Those skilled in the art will know that the number of the above-mentioned terminal devices can be more or less. For example, there can be several, dozens or hundreds, or more terminal devices. This application embodiment does not limit the number or type of terminal devices.

[0021] Server 102 can be at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. Optionally, the number of servers can be more or less, and this embodiment does not limit this. Of course, server 102 may also include other functional servers to provide more comprehensive and diversified services. In some embodiments, server 102 undertakes the main computing work, and terminal device 101 undertakes the secondary computing work; or, server 102 undertakes the secondary computing work, and terminal device 101 undertakes the main computing work; or, server 102 and terminal device 101 collaborate on computing using a distributed computing architecture. Server 102 can be connected to terminal device 101 and other terminal devices via a wireless network or a wired network. Optionally, the number of servers can be more or less, and this embodiment does not limit this.

[0022] Figure 2 This is a flowchart of an object sorting method provided according to an embodiment of this application, such as... Figure 2 As shown, this method is executed by a computer device, which can be... Figure 1 The server 102 shown includes the following steps: Step 201: Obtain the interaction behavior data of the first object in multiple business scenarios.

[0023] In this embodiment of the application, this step collects the user account interaction records generated by the first object in various business scenarios of the platform, as the basic data source for subsequent feature construction.

[0024] Optionally, the process of acquiring interactive behavior data includes three sub-steps: log reception, data clarification, and scenario classification and extraction.

[0025] First, each business module on the platform's front end reports user account behavior logs in real time. Each behavior log contains at least three core fields: the user account identifier (UserID), used to identify the unique subject of the behavior; the object identifier (ItemID), used to identify the unique object being interacted with; the business scenario identifier (SceneID), used to mark the specific scenario in which the behavior occurred, such as directly entering information in the search box, browsing a playlist page, or clicking on a theme zone; and the interaction behavior type identifier (ActionID), used to record the specific form of the interaction, such as exposure, click, play, favorite, or download.

[0026] Then, the received raw logs are preprocessed to remove duplicate logs caused by network retransmission, error logs caused by system anomalies (such as missing fields or out-of-bounds values), and false behavior logs generated by crawlers or abnormal user accounts, in order to ensure the authenticity and validity of the data.

[0027] Finally, the cleaned logs were categorized based on scenario identifiers, and all interaction records corresponding to the object identifier of the first object were extracted for each business scenario. Multiple business scenarios were divided into two main categories: search-related scenarios and non-search-related scenarios. Search-related scenarios include, but are not limited to, search scenarios where users actively input query terms, playlist-based search scenarios, and popular search term recommendation scenarios; non-search-related scenarios include, but are not limited to, browsing topic zones, browsing music library lists, using the song recognition function, and virtual radar recommendation scenarios.

[0028] Step 202: Determine the multi-dimensional features of the first object based on the interaction behavior data.

[0029] In this embodiment of the application, this step performs feature engineering on the collected interaction behavior data to construct a feature set that can comprehensively characterize the first object's performance in multiple dimensions, which serves as the input for the subsequent ranking model.

[0030] Optionally, the multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data. Accordingly, the process of determining multi-dimensional features includes three sub-steps: determining numerical features, proportional features, and time dimension features.

[0031] First, the process of determining numerical features involves extracting numerical metrics from the interaction behavior data. These numerical metrics are statistical quantities with additivity and absolute dimensions, including but not limited to the number of times the first object was played, favorited, downloaded, and shared in various business scenarios.

[0032] Secondly, the process of determining proportional characteristics involves extracting proportional indicators from interaction behavior data. Proportional indicators are relative statistical quantities that reflect the efficiency of user interaction conversion, including but not limited to the click-through rate (the ratio of clicks to impressions), playback conversion rate (the ratio of playbacks to clicks), and collection rate (the ratio of collections to playbacks) of the first object in various business scenarios.

[0033] Finally, the process of determining the time dimension features involves extracting statistical values ​​at least two different time granularities from the interaction behavior data. Time granularity includes long-cycle time granularity and short-cycle time granularity.

[0034] Step 203: Fuse the multi-dimensional features with the original associated features of the first object to obtain the fused features.

[0035] In this embodiment, this step integrates the multi-dimensional features determined in the previous steps with the pre-existing original association features to form a comprehensive feature representation (i.e., fused feature) containing multi-source information, which serves as the input to the subsequent ranking model. The original association features are used to indicate the association relationship between the first object and at least one search term.

[0036] Optionally, the process of determining the fusion features includes the following sub-steps.

[0037] First, obtain the original association features. This involves acquiring the original association features between the first object and at least one search term. These original association features are used to quantify the degree of matching and relevance between the search term and the first object. For example, the original association features can be search term and object features. This is achieved by mapping the search term and the first object to the same semantic space, calculating the similarity between their vector representations, and capturing associations that are literally mismatched but semantically related.

[0038] Secondly, feature fusion processing. This involves concatenating the multi-dimensional feature vector with the original associated feature vector along the feature dimension to generate a new, higher-dimensional feature vector. This fused feature simultaneously includes information on the interactive behavior of the first object in multiple business scenarios (from the multi-dimensional features) and information on the association between the first object and the current search term (from the original associated features), achieving complementarity and integration of cross-source information.

[0039] Step 204: Input the fused features into the multi-objective ranking model to obtain the ranking score of the first object, and rank the first object according to the ranking score.

[0040] In this embodiment of the application, this step involves inputting fused features into a multi-objective ranking model for inference, obtaining a comprehensive performance score of the first object on multiple optimization objectives, and ranking the candidate object set based on the score to generate the final ranking result. The multi-objective ranking model can be, for example, MMOE (Multi-gate Mixture-of-Experts).

[0041] Optionally, the process of determining the sort score and performing the sort includes the following sub-steps.

[0042] First, the multi-objective ranking model is input and computed. The fused features are input into the multi-objective ranking model. This model employs a multi-task learning architecture, capable of simultaneously optimizing at least two core ranking objectives. Internally, the fused features first undergo feature transformation and abstraction through a shared representation layer, extracting a general hidden layer representation applicable to multiple tasks. Subsequently, for each objective task, the model further transforms the shared representation through a specific task output layer, calculating the predicted value for each objective. The predicted values ​​for the objectives include, but are not limited to, click-through rate (CTR, the probability of a user clicking the first object) and exposure-to-conversion rate (CTCVR, the probability of a user clicking to complete a conversion after the object is exposed), reflecting the first object's attractiveness to the user and its ability to encourage deeper user interaction, respectively.

[0043] Secondly, a ranking score is generated. That is, based on the predicted values ​​of multiple objectives output by the multi-objective ranking model, the final ranking score of the first object in this ranking is calculated. This ranking score comprehensively reflects the overall performance of the first object across multiple objective dimensions; a higher score indicates that the first object better meets the user's overall expectations.

[0044] Finally, the ranking of the candidate objects is output. That is, after obtaining the ranking scores of all candidate objects (including the first object), the candidate object set is sorted in descending order of score to generate the final ranked list. This ranked list is returned to the front-end (e.g., the client) for display. The front-end presents the first object to the user according to its priority based on the ranking results.

[0045] This application provides an object ranking scheme. By acquiring interaction behavior data of a first object in multiple business scenarios, it determines multi-dimensional features including numerical features, proportional features, and multi-temporal granular features. These features are then fused with the original association features and input into a multi-objective ranking model. The multi-temporal granular features simultaneously reflect the medium-to-long-term trend and real-time changes in object popularity, enhancing the timeliness of the ranking. The multi-objective ranking model comprehensively optimizes multiple objectives, making the ranking results more consistent with users' actual needs, thus improving the overall quality and stability of the ranking effect. This scheme effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the ranking results.

[0046] The above Figure 2 The diagram shows the main flow of an object sorting method provided in this application. The process of obtaining interactive behavior data in this object sorting scheme will be further explained below. Figure 3 This is a flowchart illustrating another object sorting method provided according to an embodiment of this application. The method is executed by a computer device, such as... Figure 3 As shown, the method includes: Step 301: Receive log data reported by the client.

[0047] In this embodiment of the application, this step obtains the raw log data generated and reported by the client during the interaction between the user account and the first object, which serves as the basic data source for subsequent feature construction.

[0048] Optionally, the log data includes the user account identifier, the object identifier of the first object, the scenario identifier of the business scenario, and the behavior type of the interaction.

[0049] First, when a user account performs various operations within the platform where the first object resides, the client monitors the user account's interactive behavior in real time and generates a structured log record for each interaction. This client includes, but is not limited to, mobile applications, web applications, and desktop applications. The log data is encapsulated according to a preset data format and reported to the server's log collection system via network transmission protocols.

[0050] Secondly, each log entry contains at least four core identifier fields: UserID (user account identifier), used to uniquely identify the user subject that generated the interaction; ItemID (object identifier of the first object), used to uniquely identify the target object that the user interacted with, which can be a content entity such as a song, video, product, or news article; SceneID (scene identifier of the business scenario), used to mark the specific business scenario in which the interaction occurred, including search scenarios, playlist scenarios, theme zone scenarios, music library scenarios, etc.; and ActionType (action type identifier of the interaction), used to record the specific type of interaction action between the user and the first object, including exposure, click, play, favorite, download, share, skip, etc.

[0051] Finally, the server's log collection system performs preliminary format validation on the received log data to ensure that each log entry contains all required fields and that the field formats conform to specifications. Log data that passes validation can then be written to the original log storage system in chronological order as input for subsequent data processing.

[0052] Step 302: Clean the log data and remove abnormal and duplicate logs.

[0053] In this embodiment of the application, this step is to preprocess the received raw log data, filter out invalid and interfering data, and ensure the data quality of subsequent feature statistics.

[0054] First, the received raw log data undergoes integrity and validity checks to identify and remove various abnormal logs. Abnormal logs include at least one of the following categories: missing field logs, i.e., log data lacking any of the required fields such as user identifier, object identifier, scene identifier, or behavior type identifier; incorrect field format logs, i.e., field values ​​in log data do not conform to preset data types or format specifications, such as identifier fields containing illegal characters or numeric fields exceeding reasonable ranges; abnormal business logic logs, i.e., behavior sequences in log data that do not conform to normal user behavior logic, such as reporting click behavior when no exposure occurred, or reporting playback behavior when no click occurred; and abnormal system logs, i.e., error logs caused by system errors or network failures, such as abnormal timestamps or duplicate reporting of invalid heartbeat data.

[0055] Secondly, the log data undergoes deduplication to identify and remove duplicate logs generated due to network retransmissions, system retries, and other reasons. Duplicate log identification can employ a multi-field joint matching method. For example, the user account identifier, object identifier, scene identifier, behavior type identifier, and behavior occurrence timestamp can be combined as a unique identifier for each log entry. Logs with the same unique identifier are counted, retaining the first received log record or the log record with the earliest timestamp, and removing subsequent duplicate reports of the same log. For consecutive identical behaviors of the same user account on the same object in the same scene within a short period, a preset time interval threshold is used for judgment, and abnormal duplicate logs exceeding the normal behavior frequency are removed.

[0056] Finally, the valid log data, after anomaly and duplicate removal, is organized and output according to a preset data format to form a cleaned log dataset. The cleaned log dataset retains complete core field information, and each log entry is a valid and unique record of interaction behavior, serving as input data for subsequent scenario-based classification and feature statistics.

[0057] Step 303: Classify the cleaned log data according to the scene identifier, and extract the interaction behavior records corresponding to the object identifier of the first object in each business scene to obtain the interaction behavior data.

[0058] In this embodiment of the application, the purpose of this step is to divide the cleaned log data according to the business scenario dimension and extract all interaction records related to the first object to form an interaction behavior dataset organized by scenario.

[0059] Optionally, firstly, the scenario identifier field in the cleaned log data is read, and the log data is divided into corresponding business scenario categories according to preset scenario classification rules. These business scenarios include at least one of the following: Active search scenario (where a user actively enters a query in the search box and interacts with the user); Playlist search scenario (where a user searches for or browses songs within a playlist on the playlist page and interacts with the user); Popular search scenario (where a user clicks on a popular search term recommendation, enters the search results page, and interacts with the user); Theme zone scenario (where a user enters a specific theme or activity zone on the platform and interacts with the user); Music library scenario (where a user browses categorized playlists or song lists on the platform's music library page and interacts with the user); and Song recognition scenario (where a user identifies a song using the song recognition function and interacts with the user).

[0060] Secondly, within each categorized business scenario, using the object identifier of the first object as the query key, extract all interaction records associated with that object identifier from the scenario's log data. Optionally, this extraction process can employ key-value matching, traversing the log dataset after scenario categorization and filtering out all log entries whose object identifier field value matches the object identifier of the target first object.

[0061] Finally, the extracted interaction behavior records are stored according to business scenarios, forming an interaction behavior dataset with business scenario as the dimension and the first object as the primary key. In this dataset, each business scenario corresponds to a set of interaction behavior records. This set of records contains all historical interaction behaviors of the first object in that scenario, and each record includes information such as user account identifier, behavior type, and behavior timestamp.

[0062] Step 304: Determine the multi-dimensional features of the first object based on the interaction behavior data.

[0063] In the embodiments of this application, the multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data.

[0064] This step is the same as step 202 above, see step 202 above, and will not be repeated here.

[0065] Step 305: Fuse the multi-dimensional features with the original association features of the first object to obtain the fused features.

[0066] In this embodiment of the application, the original association feature is used to indicate the association relationship between the first object and at least one search term.

[0067] This step is the same as step 203 above, see step 203 above, and will not be repeated here.

[0068] Step 306: Input the fused features into the multi-objective ranking model to obtain the ranking score of the first object, and rank the first object according to the ranking score.

[0069] In this embodiment of the application, this step is the same as step 204 above. Please refer to step 204 above, and it will not be repeated here.

[0070] This application provides an object ranking scheme. By acquiring interaction behavior data of a first object in multiple business scenarios, it determines multi-dimensional features including numerical features, proportional features, and multi-temporal granular features. These features are then fused with the original association features and input into a multi-objective ranking model. The multi-temporal granular features simultaneously reflect the medium-to-long-term trend and real-time changes in object popularity, enhancing the timeliness of the ranking. The multi-objective ranking model comprehensively optimizes multiple objectives, making the ranking results more consistent with users' actual needs, thus improving the overall quality and stability of the ranking effect. This scheme effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the ranking results.

[0071] The above Figure 2 The diagram shows the main flow of an object sorting method provided in this application. The process of determining multidimensional features in this object sorting scheme will be further explained below. Figure 4 This is a flowchart illustrating another object sorting method provided according to an embodiment of this application. The method is executed by a computer device, such as... Figure 4 As shown, the method includes: Step 401: Obtain the interaction behavior data of the first object in multiple business scenarios.

[0072] In the embodiments of this application, multiple business scenarios include search-related scenarios and non-search-related scenarios.

[0073] This step is the same as step 201 above, see step 201 above, and will not be repeated here.

[0074] Step 402: Extract numerical and proportional indicators from the interaction behavior data.

[0075] In this embodiment of the application, this step extracts indicators from the acquired interactive behavior data, classifies and statistically analyzes the indicators according to their dimensional attributes, and obtains raw numerical indicators and raw proportional indicators, which serve as the basic input for subsequent feature processing.

[0076] First, the behavior type identifiers in the interaction behavior data are statistically counted to extract numerical indicators with absolute dimensions. This extraction process, for each business scenario's interaction behavior records, uses the object identifier of the first object as the grouping key to aggregate and count each behavior type separately, obtaining the numerical indicators of the first object in each business scenario. These numerical indicators include at least one of the following: playback count (the number of times the first object is fully or partially played by a user account); favorites count (the number of times a user account adds the first object to their favorites); download count (the number of times a user account downloads the first object to their local device); sharing count (the number of times a user account shares the first object with others through social channels); and exposure count (the number of times the first object is displayed in the user interface).

[0077] Secondly, based on the statistical results of numerical indicators, proportional indicators with relative dimensions are extracted by calculating the ratios between indicators. These proportional indicators reflect the conversion efficiency of user accounts at different behavioral stages. These proportional indicators include at least one of the following categories: Click-through rate (CTR), which is the ratio of clicks to impressions of the first object, reflecting the first object's ability to attract user clicks during display; Playback conversion rate (PCC), which is the ratio of plays to clicks of the first object, reflecting the proportion of user clicks that actually result in playback; Favorites conversion rate (Fountain collection conversion rate), which is the ratio of Favorites to plays of the first object, reflecting the proportion of user intentions to Favorite after playback; Download conversion rate (Download conversion rate), which is the ratio of downloads to plays of the first object, reflecting the proportion of user downloads after playback; and Completion rate (Complete playback rate), which is the ratio of the number of times the first object was fully played to the total number of plays, reflecting the continued attractiveness of the first object to the user.

[0078] Finally, the extracted numerical and proportional indicators are initially organized according to business scenarios and time attributes to form the original indicator dataset. In this original indicator dataset, each first object corresponds to a set of numerical indicator values ​​and a set of proportional indicator values, which serve as input data for subsequent extreme value suppression processing, small sample correction processing, and time dimension feature construction.

[0079] Step 403: Perform extreme value suppression processing on the numerical index to obtain numerical features.

[0080] In this embodiment of the application, the purpose of this step is to perform extreme value suppression transformation on the extracted original numerical indicators to eliminate the interference of extreme values ​​on subsequent model training, so that numerical indicators of different magnitudes are in a comparable feature space.

[0081] Alternatively, extremum suppression processing may include any of the following implementations.

[0082] The first implementation method is an extremum suppression processing method based on logarithmic transformation.

[0083] Correspondingly, extreme value suppression is performed on numerical indicators to obtain numerical characteristics, including: A preset constant is added to the original value of the numerical indicator to obtain a first intermediate value. This preset constant is used to avoid logarithmic calculation anomalies caused by the original value being zero. Then, the common logarithm of the first intermediate value is taken to obtain a second intermediate value. The logarithmic transformation compresses the absolute differences in values, mapping the exponentially growing original value to a linearly growing transformed value. Finally, the second intermediate value is normalized so that the transformed numerical feature falls within a preset numerical range, thus obtaining the numerical feature.

[0084] For example, the calculation formula is: .

[0085] in, The original value of the numerical indicator is given. Adding 3 is to avoid abnormal logarithmic calculations when the original data is 0. Dividing by log10 and 10.0 are to normalize the results, ensuring that the feature value is within a reasonable range. For example, when the original value increases from 1000 to 10000, the transformed value only increases from about 3.00 to about 4.00, significantly compressing the extreme value differences.

[0086] The second implementation method is an extremum suppression processing method based on square root transformation.

[0087] Correspondingly, extreme value suppression is performed on numerical indicators to obtain numerical characteristics, including: A preset constant is added to the original value of the numerical indicator to obtain a first intermediate value. This preset constant is used to avoid abnormal square root calculations caused by the original value being zero. Then, the square root of the first intermediate value is taken to obtain the numerical feature. The square root transformation can perform non-linear compression on the original value, but its compression strength is lower than that of the logarithmic transformation, making it suitable for scenarios where extreme value phenomena are relatively mild. For example, the numerical feature is... When the original value increases from 100 to 400, the transformed value increases from about 10.05 to about 20.02, and the extreme value difference is suppressed to a certain extent.

[0088] The third implementation method is an extreme value suppression processing method based on normalization of minimum and maximum values.

[0089] Correspondingly, extreme value suppression is performed on numerical indicators to obtain numerical characteristics, including: Obtain the minimum and maximum values ​​of the dataset containing the numerical metric. These values ​​reflect the distribution range of the metric across the entire dataset. For example, for the play count metric, calculate the minimum (min) and maximum (max) play counts for all first-order objects within the current time window. Then, calculate the difference between the original value of the numerical metric and the minimum value to obtain the first difference. Next, calculate the difference between the maximum value and the minimum value to obtain the second difference. Finally, divide the first difference by the second difference to obtain the numerical feature. This method linearly maps the original numerical values ​​to the interval [0,1], preserving the relative distribution of the original data while eliminating dimensions. For example, if a song has 500 plays, the minimum number of plays in the entire dataset is 10, and the maximum number of plays is 1000, then its normalized numerical characteristic is (500-10) / (1000-10)≈0.495.

[0090] Step 404: Perform small sample correction on the proportional indicators to obtain proportional features.

[0091] In this embodiment of the application, the purpose of this step is to perform small-sample correction on the extracted original proportional indicators, eliminate statistical bias caused by insufficient sample size, and make proportional indicators under different sample sizes comparable and stable.

[0092] Alternatively, small sample correction processing includes any of the following implementations.

[0093] The first implementation method is a correction process based on Wilson interval smoothing.

[0094] Accordingly, proportional indicators undergo small-sample correction to obtain proportional characteristics, including: Obtain the raw value of the proportional metric, which is determined based on the ratio of the number of clicks to the number of plays for a first object within a preset time window. For example, for the exposure conversion rate metric, if the preset time window is 1 day, then the raw click conversion rate cvr_1day = number of clicks / number of plays. If a song is played 1000 times in 1 day, and 50 of those times are clicked, then the raw click conversion rate is 0.05.

[0095] Then, a smoothing coefficient is determined based on the total number of plays of the first object within a reference time window. The reference time window is used to evaluate the sample size; for example, a 7-day reference time window is used, meaning the smoothing coefficient z is determined based on the interval where the song's total plays over the last 7 days (play_7day).

[0096] Finally, based on the original values ​​of the proportional indicators, the total number of plays, and the smoothing coefficient, the smoothed proportional characteristics are calculated using the Wilson interval lower limit formula.

[0097] Wilson's formula for the lower limit of an interval is: .

[0098] Where p is the original proportional index value, which can also be represented as cvr_1day. n is the total playback volume within the preset time window, which can also be represented as target_play_1day. z is the smoothing coefficient determined based on the total playback volume within the reference time window.

[0099] For example, if the original click-through rate of a song is p=0.05, the daily play count is n=1000, and z=15 is determined based on its total play count over 7 days, then the smoothed click-through rate is approximately 0.048, which is a slight correction to the conservative estimate.

[0100] The second implementation method is the Wilson correction method based on the hierarchical smoothing coefficient.

[0101] Accordingly, based on the total playback volume of the first object within the reference time window, a smoothing coefficient is determined, including: Based on the Wilson smoothing described above, the specific value of the smoothing coefficient is determined according to the total playback volume of the first object within the reference time window, following a layered rule: When the total number of plays is less than the first threshold, the smoothing coefficient is set to the first value. At this point, the sample size is extremely small, so a strong smoothing level is applied.

[0102] When the total number of plays is greater than or equal to the first threshold but less than the second threshold, the smoothing coefficient is determined to be the second value. At this point, the sample size is small, so a medium smoothing level is used.

[0103] When the total number of plays is greater than or equal to the second threshold but less than the third threshold, the smoothing coefficient is set to the third value. Since the sample size is large at this point, a weak smoothing level is used.

[0104] When the total number of plays is greater than or equal to the third threshold, the smoothing coefficient is set to the fourth value. At this point, the sample size is sufficient, and a very weak smoothing level is used to preserve the original statistical values ​​as much as possible.

[0105] Among them, the first threshold is less than the second threshold and less than the third threshold, and the first value is greater than the second value, greater than the third value, and greater than the fourth value.

[0106] For example, the first threshold is set to 1000, the second threshold to 10000, and the third threshold to 100000; the first value is set to 30, the second value to 20, the third value to 15, and the fourth value to 10. When a song's 7-day play count is 500 (less than 1000), the smoothing coefficient z is set to 30 for strong smoothing; when the 7-day play count is 5000 (between 1000 and 10000), z is set to 20 for medium smoothing; when the 7-day play count is 50000 (between 10000 and 100000), z is set to 15 for weak smoothing; and when the 7-day play count is 500000 (greater than 100000), z is set to 10 for very weak smoothing.

[0107] The third implementation method is a correction process based on Laplace smoothing.

[0108] Accordingly, the proportional indicators are corrected using a small sample to obtain proportional characteristics, including: First, obtain the numerator and denominator counts of the proportional indicator; where the numerator count is the number of successful events (such as the number of clicks) within the preset time window, and the denominator count is the total number of events (such as the number of plays) within the preset time window.

[0109] Then, determine the smoothing coefficient and the number of categories; where the smoothing coefficient α controls the smoothing intensity, and the number of categories k represents the number of possible result categories (for binary classification conversion rate indicators, k is usually taken as 2).

[0110] Then, add the molecule count to the smoothing coefficient to obtain the first sum; that is, molecule count + α.

[0111] Add the product of the smoothing coefficient and the number of categories to the denominator count to obtain the second sum; denominator count + α × k.

[0112] Divide the first sum by the second sum to obtain the smoothed proportional characteristic. That is, (numerator count + α) / (denominator count + α × k).

[0113] For example, a song might have 1 click and 2 plays on a given day, resulting in an initial click-to-conversion rate of 0.5. However, the sample size is extremely small, leading to low reliability. Using a smoothing coefficient α=1 and the number of categories k=2, the smoothed click-to-conversion rate would be (1+1) / (2+1×2)=2 / 4=0.5. If another song has 0 clicks and 0 plays on a given day, making it impossible to calculate the initial conversion rate, the smoothed click-to-conversion rate would be (0+1) / (0+2)=1 / 2=0.5, smoothing towards the prior probability of 0.5. As the sample size increases, the smoothed proportional characteristic gradually approaches the original proportional value.

[0114] Step 405: Extract at least two statistical values ​​with different time granularities from the interaction behavior data to obtain time dimension features.

[0115] In this embodiment of the application, the purpose of this step is to perform statistical analysis on the interaction behavior data according to multiple time granularities, extract features that can reflect the changing pattern of the popularity of the first object over time, so that the subsequent ranking model can simultaneously perceive the object's stable performance in the medium and long term and its short-term real-time fluctuations.

[0116] Optionally, the extraction process of time dimension features includes any of the following implementation methods.

[0117] Implementation Method 1: Statistical extraction based on a two-layer time window at the daily and hourly levels.

[0118] Accordingly, at least two statistical values ​​with different time granularities are extracted from the interaction behavior data to obtain time-dimensional features, including: First, the interaction data is statistically analyzed according to daily time windows to obtain at least one daily statistical value. Daily time windows are used to capture the stable popularity trend of the first object over a medium- to long-term time frame. Optionally, the statistical window uses days as the basic unit, including 1-day, 3-day, 7-day, and 30-day windows. For each daily window, the total number of interactions with the first object within that window is calculated, such as views, favorites, and click-through rate, resulting in a set of daily statistical values ​​reflecting the medium- to long-term trend.

[0119] Then, the interaction data is statistically analyzed according to hourly time windows to obtain at least one hourly statistical value. The hourly time window is used to capture the real-time popularity changes of the first object within a short time range. The statistical window uses hours as the basic unit and includes 1-hour, 3-hour, 6-hour, 12-hour, 24-hour, 36-hour, and 48-hour windows, etc. For each hourly window, the total number of interactions with the first object within that window is counted to obtain a set of hourly statistical values ​​reflecting real-time fluctuations.

[0120] Finally, the daily and hourly statistics are combined to obtain the time-dimensional features. For example, for a song, its time-dimensional features can be represented as [play count_1 day, play count_3 days, play count_7 days, play count_30 days, play count_1 hour, play count_3 hours, play count_6 hours, play count_12 hours, play count_24 hours, play count_36 hours, play count_48 hours]. This feature vector includes both the song's long-term popularity accumulation over the past month and its real-time popularity changes every hour over the past 48 hours.

[0121] The second implementation method is a statistical extraction method based on a multi-layer time window system.

[0122] Accordingly, at least two statistical values ​​with different time granularities are extracted from the interaction behavior data to obtain time-dimensional features, including: First, interactive behavior data is statistically analyzed according to multiple time granularities, including at least one long-term time granularity and at least one short-term time granularity. The long-term time granularity reflects the basic popularity and historical performance of the primary object, with statistical windows typically measured in days, weeks, or months. The short-term time granularity reflects the recent activity and real-time popularity of the primary object, with statistical windows typically measured in hours or minutes. The long-term and short-term time granularities constitute a multi-layered time window system of "short-term-medium-long-term." This system ensures a comprehensive depiction of the popularity changes of the primary object across different time scales. For example, for product objects in e-commerce scenarios, a multi-layered time window system can be constructed, including 1 hour, 6 hours, 24 hours (short-term), 3 days, 7 days (medium-term), 30 days, and 90 days (long-term).

[0123] Finally, the statistical values ​​corresponding to each time granularity are obtained as time dimension features. That is, at each time granularity, the numerical and proportional indicators in the interaction behavior data can be statistically analyzed separately to form a feature subset under that time granularity. The feature subsets under all time granularities are concatenated to obtain the final time dimension features.

[0124] Step 406: Fuse the multi-dimensional features with the original associated features of the first object to obtain the fused features.

[0125] In this embodiment of the application, the original association feature is used to indicate the association relationship between the first object and at least one search term.

[0126] This step is the same as step 203 above, see step 203 above, and will not be repeated here.

[0127] Step 407: Input the fused features into the multi-objective ranking model to obtain the ranking score of the first object, and rank the first object according to the ranking score.

[0128] In this embodiment of the application, this step is the same as step 203 above. Please refer to step 203 above, and it will not be repeated here.

[0129] This application provides an object ranking scheme. By acquiring interaction behavior data of a first object in multiple business scenarios, it determines multi-dimensional features including numerical features, proportional features, and multi-temporal granular features. These features are then fused with the original association features and input into a multi-objective ranking model. The multi-temporal granular features simultaneously reflect the medium-to-long-term trend and real-time changes in object popularity, enhancing the timeliness of the ranking. The multi-objective ranking model comprehensively optimizes multiple objectives, making the ranking results more consistent with users' actual needs, thus improving the overall quality and stability of the ranking effect. This scheme effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the ranking results.

[0130] The above Figure 2 The diagram shows the main flow of an object sorting method provided in this application. The processing of the multi-objective sorting model in this object sorting scheme will be further explained below. Figure 5 This is a flowchart illustrating another object sorting method provided according to an embodiment of this application. The method is executed by a computer device, such as... Figure 5 As shown, the method includes: Step 501: Obtain the interaction behavior data of the first object in multiple business scenarios.

[0131] In the embodiments of this application, multiple business scenarios include search-related scenarios and non-search-related scenarios.

[0132] This step is the same as step 201 above, see step 201 above, and will not be repeated here.

[0133] Step 502: Determine the multi-dimensional features of the first object based on the interaction behavior data.

[0134] In the embodiments of this application, the multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data.

[0135] This step is the same as step 202 above, see step 202 above, and will not be repeated here.

[0136] Step 503: Fuse the multi-dimensional features with the original associated features of the first object to obtain the fused features.

[0137] In this embodiment of the application, the original association feature is used to indicate the association relationship between the first object and at least one search term.

[0138] This step is the same as step 203 above, see step 203 above, and will not be repeated here.

[0139] Step 504: Input the fused features into the multi-objective ranking model, which includes multiple expert networks and multiple gating networks.

[0140] In this embodiment of the application, the fused features are used as input data and fed into a multi-objective ranking model with a multi-expert and multi-gated architecture. Through the collaborative computation of the expert network and the gated network, a task-specific hidden layer representation is provided for subsequent multi-objective prediction.

[0141] First, the fused feature vector is used as input layer data for the multi-objective ranking model. Optionally, the fused feature vector contains all the information of the multi-dimensional features of the first object and the original associated features. This fused feature vector is simultaneously passed to multiple expert networks at the bottom layer of the model, allowing each expert network to perform independent feature transformations.

[0142] Step 505: Transform the fused features through each expert network to obtain multiple expert outputs.

[0143] In this embodiment, multiple expert networks receive the same fused feature input in parallel, and each expert network is an independent nonlinear transformation unit.

[0144] For example, if the model contains k expert networks, the output of the i-th expert network is calculated as f_i(x) = activation(W_i·x + b_i), where W_i is the weight matrix of the i-th expert network, b_i is the bias vector, and activation is a non-linear activation function (such as ReLU, tanh, etc.). Each expert network abstracts and transforms the input features from different perspectives, learning different hidden layer representations of the input features. If the output dimensions of the k expert networks are the same, denoted as h, then the outputs of all expert networks constitute a k×h expert output matrix E = [f_1(x), f_2(x), ..., f_k(x)]^T.

[0145] Step 506: For each target, the weights of each expert output are generated based on the fusion features through the corresponding gating network, and the weighted sum of each expert output is obtained to obtain the hidden layer representation of the target.

[0146] In this embodiment, a corresponding gating network is set up for each objective task that the multi-objective ranking model needs to optimize. Assuming the model needs to optimize m objectives simultaneously (e.g., click-through rate (CTR) and click-through rate conversion rate (CTCVR), then m=2), the model contains m gating networks. Each gating network takes a fused feature x as input and generates a k-dimensional weight vector through a softmax function, representing the importance of the outputs of the k expert networks for that task. The output of the j-th gating network is calculated as g_j(x)=softmax(V_j·x+c_j), where V_j is the weight matrix of the j-th gating network, c_j is the bias vector, and the softmax function ensures that the sum of the k output weights is 1. The output of the gating network is g_j(x)=[α_{j1},α_{j2},...,α_{jk}], where α_{ji} represents the contribution weight of the ith expert network for the j-th task.

[0147] For each target task, the weight vector output by the corresponding gating network is weighted and summed with the outputs of all expert networks to obtain the task-specific hidden layer representation. The hidden layer representation of the j-th task is calculated as t_j = Σ_{i=1}^{k}α_{ji}·f_i(x), which is a weighted combination of the outputs of each expert network. Since the gating networks of different tasks generate different weight distributions, the final hidden layer representation t_j obtained for different tasks will also be different, realizing a multi-objective learning mechanism of "shared expert networks, task-specific combinations". For example, for the CTR task, its gating network may focus more on the output of the expert network that captures the user's click tendency; for the CTCVR task, its gating network may focus more on the output of the expert network that captures the conversion intention.

[0148] Step 507: Input the hidden layer representation of each target into the corresponding target prediction network to obtain the predicted value of each target.

[0149] In this embodiment, firstly, the hidden layer representation vector corresponding to each target task is obtained. This hidden layer representation t_j has been fused with feature information extracted from different perspectives by multiple expert networks through a weighted combination mechanism of a gating network, and has been adaptively adjusted for the task characteristics of the j-th target.

[0150] Then, an independent target prediction network is set up for each target task. This target prediction network typically consists of a single-layer or multi-layer fully connected network, with the final output layer employing an activation function suitable for that target task. The j-th target prediction network takes the corresponding hidden layer representation t_j as input and calculates the predicted value of the j-th target through forward propagation. The forward propagation calculation process is represented as p_j = σ_j(W_j·t_j + b_j), where W_j is the weight matrix of the j-th target prediction network, b_j is the bias vector, and σ_j is the activation function of the output layer corresponding to the j-th target. For binary classification targets (such as click-through rate (CTR), predicting whether a user will click), the output layer activation function is usually the Sigmoid function, mapping the network output to the (0,1) interval, representing the probability of a click; for regression targets (such as playback duration prediction), the output layer activation function can be a linear function or the ReLU function, directly outputting the predicted value.

[0151] Finally, all m target prediction networks are computed in parallel, and the predicted values ​​for each target task are output simultaneously, forming a multi-target prediction result set.

[0152] For example, in an optimization scenario with two objectives, CTR and CTCVR, the model simultaneously outputs two predicted values: p_CTR represents the predicted probability that the first object will be clicked by the user, and p_CTCVR represents the predicted probability that after the first object is exposed, the user will click on the first object to further complete a conversion behavior (such as purchase, download, full playback, etc.). The two predicted values ​​are calculated based on the same fused feature input through a shared expert network and task-specific gating and prediction networks, achieving joint optimization of multiple objectives within a multi-task learning framework.

[0153] To make the processing of the multi-objective ranking model easier to understand, see [link to relevant documentation]. Figure 6 As shown, Figure 6 This is a flowchart illustrating the processing of a multi-objective ranking model according to an embodiment of this application. Figure 6 The process includes: 601. Inputting client logs. 602. Data cleaning and statistics. 603. Storing in a database, such as Redis. Alternatively, databases supporting high-concurrency reads, such as Memcached or TiDB, can be used to replace Redis, ensuring high performance and high availability for feature storage. 604. Obtaining global features from the database, i.e., the multi-dimensional features mentioned above. 605. Obtaining search domain features from the database, i.e., the original association features mentioned above. 606. Feature input. 607. Inputting into the expert network. 608. Inputting into the gating network. 609. Inputting into the CTR prediction network. 610. Inputting into the CVR prediction network. 611. Outputting p_CTR. 612. Outputting p_CTCVR.

[0154] Step 508: Calculate the ranking score of the first object based on the predicted values ​​of each target.

[0155] In this embodiment of the application, this step merges the multiple target prediction values ​​output by the multi-objective ranking model into a single comprehensive ranking score.

[0156] First, obtain the predicted values ​​of the first object for each optimization objective. Assuming the model optimizes m objectives simultaneously, and the predicted value of the j-th objective is denoted as p_j, then the set of multi-objective predicted values ​​for the first object is {p_1, p_2, ..., p_m}.

[0157] For example, in a scenario where two objectives are being optimized, click-through rate (CTR) and click-through conversion rate (CTCVR), the set of predicted values ​​for the first object is {p_CTR, p_CTCVR}, where p_CTR represents the predicted probability that a user will click on the first object, with a value between 0 and 1; p_CTCVR represents the predicted probability that a user will complete a conversion after clicking, also with a value between 0 and 1.

[0158] Then, based on the preset weight configuration, the predicted values ​​of multiple targets are weighted and fused to calculate the comprehensive ranking score of the first object. The weight configuration reflects the importance of different business objectives in the final ranking decision and is preset according to the platform's business strategy. Let the weight of the j-th objective be w_j, satisfying w_j≥0 and typically ∑{j=1}^{m}w_j=1. Then, the comprehensive ranking score S of the first object is calculated as the weighted sum of the predicted values ​​of each objective: S=∑{j=1}^{m}w_j·p_j.

[0159] For example, if the business prioritizes user click behavior, the CTR weight can be set to 0.7 and the CTCVR weight to 0.3; if the business prioritizes the final conversion effect, the CTR weight can be set to 0.4 and the CTCVR weight to 0.6. For a given object, if its p_CTR=0.8 and p_CTCVR=0.3, with a CTR weight of 0.7 and a CTCVR weight of 0.3, its overall ranking score S=0.7×0.8+0.3×0.3=0.56+0.09=0.65.

[0160] Step 509: Sort the first object according to the sorting score.

[0161] In this embodiment, after obtaining the comprehensive ranking score of all candidate objects, the candidate object set is sorted in descending order of score. During the sorting process, the object with the highest score is placed at the beginning of the sorted list, followed by the object with the second highest score, and so on, to generate the final sorted result list.

[0162] For example, given a candidate song set {A, B, C, D}, if the calculated comprehensive ranking scores for each song are S_A=0.65, S_B=0.82, S_C=0.41, and S_D=0.73, then the list sorted from highest to lowest score is [B, D, A, C]. This ranking list is returned to the client, which then presents the first item to the user in the order listed, thus optimizing the display of search results.

[0163] This application provides an object ranking scheme. By acquiring interaction behavior data of a first object in multiple business scenarios, it determines multi-dimensional features including numerical features, proportional features, and multi-temporal granular features. These features are then fused with the original association features and input into a multi-objective ranking model. The multi-temporal granular features simultaneously reflect the medium-to-long-term trend and real-time changes in object popularity, enhancing the timeliness of the ranking. The multi-objective ranking model comprehensively optimizes multiple objectives, making the ranking results more consistent with users' actual needs, thus improving the overall quality and stability of the ranking effect. This scheme effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the ranking results.

[0164] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user data. These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their data is being collected. This ensures that the application only begins the steps for collecting user data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without user confirmation), the steps for collecting user data end, meaning no user data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0165] It should be noted that the order of the method steps provided in the embodiments of this application can be appropriately adjusted, and the steps can also be added or removed as appropriate. Any method variations that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application, and therefore will not be elaborated further.

[0166] Figure 7 This is a schematic diagram of the structure of an object sorting device provided according to an embodiment of this application. For example... Figure 7 As shown, the device includes: an acquisition module 701, a determination module 702, a fusion module 703, and a sorting module 704.

[0167] The acquisition module 701 is used to acquire the interaction behavior data of the first object in multiple business scenarios, including search-related scenarios and non-search-related scenarios. The determination module 702 is used to determine the multi-dimensional features of the first object based on the interaction behavior data. The multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data. The fusion module 703 is used to fuse multi-dimensional features with the original association features of the first object to obtain fused features. The original association features are used to indicate the association relationship between the first object and at least one search term. The sorting module 704 is used to input the fused features into the multi-objective sorting model, obtain the sorting score of the first object, and sort the first object according to the sorting score.

[0168] In some embodiments, the acquisition module 701 is used to receive log data reported by the client. The log data includes the account identifier of the user account, the object identifier of the first object, the scenario identifier of the business scenario, and the behavior type of the interaction behavior. The log data is cleaned to remove abnormal logs and duplicate logs. The cleaned log data is classified according to the scenario identifier, and the interaction behavior records corresponding to the object identifier of the first object under each business scenario are extracted to obtain the interaction behavior data.

[0169] In some embodiments, the determining module 702 is used to extract numerical indicators and proportional indicators from the interaction behavior data; perform extreme value suppression processing on the numerical indicators to obtain numerical features; perform small sample correction processing on the proportional indicators to obtain proportional features; and extract at least two statistical values ​​with different time granularities from the interaction behavior data to obtain time dimension features.

[0170] In some embodiments, the determining module 702 is used to add a preset constant to the original value of the numerical index to obtain a first intermediate value; take the common logarithm of the first intermediate value to obtain a second intermediate value; and normalize the second intermediate value to obtain a numerical feature.

[0171] In some embodiments, the determining module 702 is used to add a preset constant to the original value of the numerical index to obtain a first intermediate value; and to take the square root of the first intermediate value to obtain a numerical feature.

[0172] In some embodiments, the determining module 702 is used to obtain the minimum and maximum values ​​of the dataset containing the numerical indicator; calculate the difference between the original value of the numerical indicator and the minimum value to obtain a first difference; calculate the difference between the maximum value and the minimum value to obtain a second difference; and divide the first difference by the second difference to obtain the numerical feature.

[0173] In some embodiments, the determining module 702 is used to obtain the original value of the proportional indicator, the original value being determined based on the ratio of the number of clicks to the number of plays of the first object within a preset time window; to determine the smoothing coefficient based on the total number of plays of the first object within a reference time window; and to calculate the smoothed proportional feature based on the original value of the proportional indicator, the total number of plays, and the smoothing coefficient using the Wilson interval lower limit formula.

[0174] In some embodiments, the determining module 702 is configured to determine a smoothing coefficient as a first value when the total playback volume is less than a first threshold; determine a smoothing coefficient as a second value when the total playback volume is greater than or equal to the first threshold and less than a second threshold; determine a smoothing coefficient as a third value when the total playback volume is greater than or equal to the second threshold and less than a third threshold; and determine a smoothing coefficient as a fourth value when the total playback volume is greater than or equal to the third threshold; wherein the first threshold is less than the second threshold and less than the third threshold, and the first value is greater than the second value, greater than the third value, and greater than the fourth value.

[0175] In some embodiments, the determining module 702 is used to obtain the numerator count and denominator count of the proportional index; determine the smoothing coefficient and the number of categories; add the numerator count and the smoothing coefficient to obtain a first sum; add the product of the smoothing coefficient and the number of categories to the denominator count to obtain a second sum; and divide the first sum by the second sum to obtain the smoothed proportional feature.

[0176] In some embodiments, the determining module 702 is used to perform statistics on the interaction behavior data according to the daily time window to obtain at least one daily statistical value; to perform statistics on the interaction behavior data according to the hourly time window to obtain at least one hourly statistical value; and to combine the daily statistical value and the hourly statistical value to obtain the time dimension feature.

[0177] In some embodiments, the determining module 702 is used to perform statistics on interactive behavior data according to multiple time granularities, the multiple time granularities including at least one long-cycle time granularity and at least one short-cycle time granularity, the long-cycle time granularity and the short-cycle time granularity constitute a multi-layer time window system of "short-term-medium-long-term"; and to obtain the statistical value corresponding to each time granularity as a time dimension feature.

[0178] In some embodiments, the ranking module 704 is used to input fused features into a multi-objective ranking model, the multi-objective ranking model including multiple expert networks and multiple gating networks; transform the fused features through each expert network to obtain multiple expert outputs; for each object, generate weights for each expert output based on the fused features through the corresponding gating network, and perform weighted summation of each expert output to obtain the hidden layer representation of the object; input the hidden layer representation of each object into the corresponding object prediction network to obtain the predicted value of each object; and calculate the ranking score of the first object based on the predicted values ​​of each object.

[0179] In some embodiments, multiple business scenarios include at least one of the following: proactive search scenario, playlist search scenario, popular search scenario, theme zone scenario, music library scenario, and music recognition scenario.

[0180] This application provides an object sorting device that acquires interaction behavior data of a first object in multiple business scenarios, determines multi-dimensional features including numerical features, proportional features, and multi-temporal granular features, and then fuses these features with the original association features before inputting them into a multi-objective sorting model. The multi-temporal granular features simultaneously reflect the medium-to-long-term trend and real-time changes in object popularity, enhancing the timeliness of the sorting. The multi-objective sorting model comprehensively optimizes multiple objectives, making the sorting results more consistent with users' actual needs, thus improving the overall quality and stability of the sorting effect. This solution effectively solves the problems of single features, data bias, and poor timeliness in existing technologies, improving the accuracy of the sorting results.

[0181] It should be noted that the object sorting device provided in the above embodiments is only an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the object sorting device and the object sorting method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0182] Embodiments of this application also provide a computer device including a processor and a memory, wherein the memory stores a computer program that is loaded and executed by the processor to implement the object sorting method provided in the above-described method embodiments.

[0183] Figure 8 This is a schematic diagram of the structure of a computer device provided according to an embodiment of this application.

[0184] Computer device 800 includes a central processing unit (CPU) 801, a system memory 804 including random access memory (RAM) 802 and read-only memory (ROM) 803, and a system bus 805 connecting the system memory 804 and the CPU 801. Computer device 800 also includes a basic input / output system (I / O system) 806 that facilitates information transfer between various devices within the computer device, and a mass storage device 807 for storing the operating system 813, application programs 814, and other program modules 815.

[0185] The basic input / output system 806 includes a display 808 for displaying information and an input device 809 for user input, such as a mouse or keyboard. Both the display 808 and the input device 809 are connected to the central processing unit 801 via an input / output controller 810 connected to the system bus 805. The basic input / output system 806 may also include the input / output controller 810 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 810 also provides output to a display screen, printer, or other types of output devices.

[0186] Mass storage device 807 is connected to central processing unit 801 via a mass storage controller (not shown) connected to system bus 805. Mass storage device 807 and its associated computer-readable storage media provide non-volatile storage for computer device 800. That is, mass storage device 807 may include computer-readable storage media (not shown) such as hard disk or compact disc read-only memory (CD-ROM) drive.

[0187] Without loss of generality, computer-readable storage media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technique for storing information such as computer-readable storage instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically-erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage devices, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 804 and mass storage device 807 described above can be collectively referred to as memory.

[0188] The memory stores one or more programs, which are configured to be executed by one or more central processing units 801. The one or more programs contain instructions for implementing the above method embodiments, and the central processing unit 801 executes the one or more programs to implement the methods provided by the above method embodiments.

[0189] According to various embodiments of this application, the computer device 800 can also be connected to a remote computer device on a network, such as the Internet, for operation. That is, the computer device 800 can be connected to a network 812 via a network interface unit 811 connected to the system bus 805, or the network interface unit 811 can be used to connect to other types of networks or remote computer device systems (not shown).

[0190] The memory also includes one or more programs stored in the memory, and the one or more programs include steps performed by a computer device in the methods provided in the embodiments of this application.

[0191] This application also provides a computer-readable storage medium storing a computer program that is loaded and executed by a processor to implement the object sorting method provided in the above-described method embodiments.

[0192] This application also provides a computer program product comprising a computer program executed by a processor to implement the object sorting methods provided in the above-described method embodiments.

[0193] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0194] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent switching, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An object sorting method, characterized in that, The method includes: Acquire interactive behavior data of a first object in multiple business scenarios, including search-related scenarios and non-search-related scenarios; Based on the interaction behavior data, the multi-dimensional features of the first object are determined. The multi-dimensional features include at least one of the following: numerical features representing numerical indicators in the interaction behavior data, proportional features representing proportional indicators in the interaction behavior data, and time dimension features representing at least two different time granularities in the interaction behavior data. The multi-dimensional features are fused with the original association features of the first object to obtain fused features, wherein the original association features are used to indicate the association relationship between the first object and at least one search term; The fused features are input into a multi-objective ranking model to obtain a ranking score for the first object, and the first object is ranked according to the ranking score.

2. The method according to claim 1, characterized in that, The acquisition of interaction behavior data of the first object in multiple business scenarios includes: Receive log data reported by the client, wherein the log data includes the account identifier of the user account, the object identifier of the first object, the scenario identifier of the business scenario, and the behavior type of the interaction behavior; The log data is cleaned to remove abnormal and duplicate logs; The cleaned log data is classified according to the scenario identifier, and the interaction behavior records corresponding to the object identifier of the first object in each business scenario are extracted to obtain the interaction behavior data.

3. The method according to claim 1 or 2, characterized in that, The step of determining the multi-dimensional features of the first object based on the interaction behavior data includes: Numerical and proportional indicators are extracted from the interactive behavior data; The numerical characteristics are obtained by performing extreme value suppression processing on the numerical index. The proportional indicator is subjected to small sample correction processing to obtain the proportional feature; At least two statistical values ​​with different time granularities are extracted from the interaction behavior data to obtain the time dimension features.

4. The method according to claim 3, characterized in that, The extreme value suppression processing of the numerical index to obtain the numerical features includes: A preset constant is added to the original value of the numerical index to obtain a first intermediate value; Take the common logarithm of the first intermediate value to obtain the second intermediate value; The second intermediate value is normalized to obtain the numerical feature.

5. The method according to claim 3, characterized in that, The extreme value suppression processing of the numerical index to obtain the numerical features includes: A preset constant is added to the original value of the numerical index to obtain a first intermediate value; The numerical feature is obtained by taking the square root of the first intermediate value.

6. The method according to claim 3, characterized in that, The extreme value suppression processing of the numerical index to obtain the numerical features includes: Obtain the minimum and maximum values ​​of the dataset containing the numerical indicator; Calculate the difference between the original value of the numerical index and the minimum value to obtain the first difference; Calculate the difference between the maximum value and the minimum value to obtain a second difference; Divide the first difference by the second difference to obtain the numerical feature.

7. The method according to claim 3, characterized in that, The step of performing small-sample correction processing on the proportional index to obtain the proportional feature includes: Obtain the original value of the proportional indicator, which is determined based on the ratio of the number of clicks to the number of plays of the first object within a preset time window; The smoothing coefficient is determined based on the total playback volume of the first object within the reference time window; Based on the original value of the proportional indicator, the total number of plays, and the smoothing coefficient, the smoothed proportional feature is calculated using the Wilson interval lower limit formula.

8. The method according to claim 7, characterized in that, The step of determining the smoothing coefficient based on the total playback volume of the first object within the reference time window includes: When the total playback volume is less than the first threshold, the smoothness coefficient is determined to be the first value; When the total playback volume is greater than or equal to the first threshold and less than the second threshold, the smoothing coefficient is determined to be the second value; When the total playback volume is greater than or equal to the second threshold and less than the third threshold, the smoothing coefficient is determined to be the third value; When the total playback volume is greater than or equal to the third threshold, the smoothing coefficient is determined to be the fourth value; Wherein, the first threshold is less than the second threshold is less than the third threshold, and the first value is greater than the second value is greater than the third value is greater than the fourth value.

9. The method according to claim 3, characterized in that, The step of performing small-sample correction processing on the proportional index to obtain the proportional feature includes: Obtain the numerator and denominator counts of the proportional index; Determine the smoothing coefficient and the number of categories; The molecule count is added to the smoothing coefficient to obtain a first sum. Add the product of the smoothing coefficient and the number of categories to the count in the denominator to obtain the second sum; Divide the first sum by the second sum to obtain the smoothed proportional feature.

10. The method according to claim 3, characterized in that, The step of extracting at least two statistical values ​​with different time granularities from the interaction behavior data to obtain the time dimension features includes: The interaction behavior data is statistically analyzed according to a daily time window to obtain at least one daily statistical value; The interaction behavior data is statistically analyzed according to hourly time windows to obtain at least one hourly statistical value; The time dimension feature is obtained by combining the daily statistical values ​​and the hourly statistical values.

11. The method according to claim 3, characterized in that, The step of extracting at least two statistical values ​​with different time granularities from the interaction behavior data to obtain the time dimension features includes: The interactive behavior data is statistically analyzed according to multiple time granularities, including at least one long-cycle time granularity and at least one short-cycle time granularity. The long-cycle time granularity and the short-cycle time granularity constitute a multi-layer time window system of "short-term-medium-long-term". The statistical values ​​corresponding to each time granularity are obtained and used as the time dimension features.

12. The method according to any one of claims 1 to 11, characterized in that, The step of inputting the fused features into a multi-objective ranking model to obtain the ranking score of the first object includes: The fused features are input into a multi-objective ranking model, which includes multiple expert networks and multiple gating networks. The fused features are transformed through each expert network to obtain multiple expert outputs; For each target, the weights of each expert output are generated based on the fusion features through the corresponding gating network, and the weighted sum of each expert output is obtained to obtain the hidden layer representation of the target. The hidden layer representation of each target is input into the corresponding target prediction network to obtain the predicted value of each target; Based on the predicted values ​​of each target, the ranking score of the first object is calculated.

13. The method according to any one of claims 1 to 12, characterized in that, The multiple business scenarios include at least one of the following: proactive search scenario, playlist search scenario, popular search scenario, themed zone scenario, music library scenario, and music recognition scenario.

14. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the object sorting method as described in any one of claims 1 to 13.

15. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program, which is loaded and executed by a processor to implement the object sorting method as described in any one of claims 1 to 13.

16. A computer program product, characterized in that, The computer program product includes a computer program executed by a processor to implement the object sorting method as described in any one of claims 1 to 13.

Images