Song list sorting method, computer device, readable storage medium and program product

Abstract

The present application relates to a playlist sorting method, a computer device, a readable storage medium and a program product, the method comprising: obtaining a search term input by a target user and a plurality of target playlists associated with the search term; for each of the target playlists, respectively obtaining the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions; respectively inputting the attribute characteristics, the matching degree and the historical conversion rate of each of the target playlists into a pre-trained playlist sorting model to obtain a recommendation value for each of the target playlists; sorting the plurality of target playlists according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term. The use of this method can improve the accuracy of the playlists displayed to the user.

Description

Technical Field

[0001] The present application relates to the field of search technology, and in particular to a playlist sorting method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] In recent years, the rapid development of digital music and the rise of online music platforms have left users with a vast amount of music content to choose from. In this context, how to effectively recommend personalized playlists to users has become an important topic in the research and application of music recommendation systems.

[0003] Traditional recommendation algorithms mostly rely on users' historical behavior data, such as songs they have listened to, artists they like, etc. However, users' listening habits and preferences may change, and relying solely on historical data is difficult to reflect users' current needs in a timely manner. Therefore, the recommendation accuracy is low. Summary of the invention

[0004] Based on this, it is necessary to provide a playlist sorting method, apparatus, computer device, computer-readable storage medium and computer program product to address the technical problem that the above method has low recommendation accuracy.

[0005] In a first aspect, the present application provides a method for sorting a playlist. The method comprises:

[0006] Acquire a search term input by a target user and a plurality of target playlists associated with the search term;

[0007] For each of the target playlists, the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension are obtained respectively; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions;

[0008] Inputting the attribute features, the matching degree and the historical conversion rate of each target playlist into a pre-trained playlist sorting model respectively to obtain a recommendation value for each target playlist;

[0009] The multiple target playlists are sorted according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term.

[0010] In one of the embodiments, the preset dimensions include at least one of a music market dimension, a search term dimension, and a target user group dimension to which the target user belongs. The historical conversion rate under the music market dimension reflects the popularity of the target playlist in the music market, the historical conversion rate under the search term dimension reflects the degree of matching between the target playlist and the search term, and the historical conversion rate under the target user group dimension reflects the preference of the target playlist in the target user group.

[0011] In one embodiment, obtaining the historical conversion rate of the target playlist under the preset dimension includes:

[0012] Obtain historical interaction data of the target playlist under the preset dimension;

[0013] According to the historical interaction data, the historical conversion rate of the target playlist under the preset dimension is obtained by statistics.

[0014] In one embodiment, obtaining the historical conversion rate of the target playlist under the preset dimension based on the historical interaction data includes:

[0015] Determine whether the number of successful conversions corresponding to the historical conversion rate conforms to a binomial distribution;

[0016] If it conforms to the binomial distribution, the historical interaction data of the target playlist under the preset dimension is statistically analyzed to obtain conversion parameters; the conversion parameters include the number of successful conversions and the total number of conversions;

[0017] The conversion parameters are processed by a smoothing estimation method to obtain the historical conversion rate of the target playlist under the preset dimension.

[0018] In one embodiment, the method further comprises:

[0019] If it does not conform to the binomial distribution, then the average conversion parameter of all historical playlists is calculated based on the total historical interaction data of all historical playlists under the preset dimension, and the conversion parameter of the target playlist is calculated based on the historical interaction data of the target playlist under the preset dimension;

[0020] The historical conversion rate of the target playlist under the preset dimension is calculated by using the average conversion parameter of all the historical playlists and the conversion parameter of the target playlist.

[0021] In one embodiment, the method further comprises:

[0022] If it does not conform to the binomial distribution, the number of successful conversions corresponding to the historical conversion rate is modeled as multiple binomial distribution models according to the historical interaction data of the target playlist under the preset dimension;

[0023] The multiple binomial distribution models are smoothed to obtain the historical conversion rate of the target playlist under the preset dimension.

[0024] In one embodiment, according to the historical interaction data of the target playlist under the preset dimension, the number of successful conversions corresponding to the historical conversion rate is modeled as multiple binomial distribution models, including:

[0025] Define the maximum number of successful conversions of the target playlist in a single exposure;

[0026] According to the historical interaction data, the number of successful conversions of the target playlist at each exposure is counted;

[0027] The occurrence frequency of each conversion success number that is less than the highest success number is counted respectively to construct the multiple binomial distribution models.

[0028] In one embodiment, the playlist sorting model is trained in the following manner:

[0029] Acquire sample data; the sample data includes: historical search terms of a sample user, historical playlists associated with the historical search terms, and interaction data of the sample user on the historical playlists; the historical playlists have corresponding sample tags;

[0030] Obtaining sample attribute characteristics of the historical playlist, sample matching degree with the user portrait of the sample user, and historical conversion rate under preset dimensions;

[0031] The sample attribute characteristics of the historical playlist, the sample matching degree and the historical conversion rate are used as input variables, and the sample labels are used as supervision information to train the initial playlist sorting model to obtain the pre-trained playlist sorting model.

[0032] In one embodiment, obtaining sample data includes:

[0033] Obtain the average number of playlists that can be displayed on the playlist display page, and construct a playlist sampling window according to the average number of playlists;

[0034] Determine a song list sampling window with positive samples, and perform song list sampling in the form of probability attenuation within the song list sampling window with positive samples to obtain negative samples;

[0035] The positive samples and the negative samples are combined into the sample data.

[0036] In a second aspect, the present application also provides a song list sorting device. The device comprises:

[0037] A term acquisition module, used to acquire a search term input by a target user and a plurality of target playlists associated with the search term;

[0038] A feature acquisition module is used to obtain, for each of the target playlists, the attribute features of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions;

[0039] A recommendation value generation module, used to input the attribute characteristics, the matching degree and the historical conversion rate of each target playlist into a pre-trained playlist sorting model to obtain a recommendation value for each target playlist;

[0040] A playlist sorting module is used to sort the multiple target playlists according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term.

[0041] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0042] Acquire a search term input by a target user and a plurality of target playlists associated with the search term;

[0043] For each of the target playlists, the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension are obtained respectively; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions;

[0044] Inputting the attribute features, the matching degree and the historical conversion rate of each target playlist into a pre-trained playlist sorting model respectively to obtain a recommendation value for each target playlist;

[0045] The multiple target playlists are sorted according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term.

[0046] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0047] Acquire a search term input by a target user and a plurality of target playlists associated with the search term;

[0048] For each of the target playlists, the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension are respectively obtained; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions;

[0049] Inputting the attribute features, the matching degree and the historical conversion rate of each target playlist into a pre-trained playlist sorting model respectively to obtain a recommendation value for each target playlist;

[0050] The multiple target playlists are sorted according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term.

[0051] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0052] Acquire a search term input by a target user and a plurality of target playlists associated with the search term;

[0053] For each of the target playlists, the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension are respectively obtained; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions;

[0054] Inputting the attribute features, the matching degree and the historical conversion rate of each target playlist into a pre-trained playlist sorting model respectively to obtain a recommendation value for each target playlist;

[0055] The multiple target playlists are sorted according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term.

[0056] The above-mentioned playlist sorting method, device, computer equipment, computer-readable storage medium and computer program product obtain the search terms input by the target user and multiple target playlists associated with the search terms, and respectively obtain the attribute characteristics of each target playlist, the degree of matching with the user portrait of the target user, and the historical conversion rate under the preset dimensions. The playlist recommendation model then sorts the playlist based on the attribute characteristics, matching degree and historical conversion rate of each target playlist. This method combines the degree of matching between the target playlist and the user portrait, as well as the attribute characteristics of the target playlist to sort the playlist, and can recommend playlists that meet the user's interests, achieving personalized and accurate playlist recommendations. At the same time, combined with the historical conversion rate of the playlist under the preset dimensions to sort the playlist, it can recommend playlists with higher historical conversion rates. This songlist sorting based on multi-dimensional features can improve the accuracy of the playlists displayed to the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of a flow chart of a song list sorting method in one embodiment;

[0058] Figure 2 A flowchart of a song list sorting model training step in one embodiment;

[0059] Figure 3 A schematic diagram of the principle of the training process of a playlist sorting model in one embodiment;

[0060] Figure 4 A schematic diagram of a search log association process in one embodiment;

[0061] Figure 5 A schematic diagram of a multi-source feature data stream for song list sorting in one embodiment;

[0062] Figure 6 A schematic diagram of sample collection in one embodiment;

[0063] Figure 7 It is a structural block diagram of a song list sorting device in one embodiment;

[0064] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0066] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein.

[0067] This application focuses on the field of music search and designs a sorting scheme for playlist search based on the characteristics of music business scenarios, which fully reflects the user's real preferences and feedback in sample labeling and positive and negative sample sampling; in the feature calculation link, the playlist features in different dimensions are fully utilized, and a series of feature smoothing methods are designed according to the characteristics of playlist conversion indicator data, so as to more accurately characterize the playlist performance; the data stream that is finally put online can process and generate raw data from different sources and with different update frequencies, so that the model can obtain highly timely playlist features, thereby providing more accurate, diversified and personalized playlist sorting results that meet the user's needs.

[0068] refer to Figure 1 , which is a flow chart of a playlist sorting method provided in an embodiment of the present application. This embodiment takes the method applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Among them, the terminal can be but is not limited to various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as an independent server or a server cluster composed of multiple servers. In this embodiment, the method includes the following steps:

[0069] Step S110, obtaining the search term input by the target user and multiple target playlists associated with the search term.

[0070] Among them, the search term refers to the keyword or phrase entered by the user in the music application or music search platform, which may include music style, subject, emotion, etc. For example, the search term may be "emo playlist".

[0071] In a specific implementation, the terminal can respond to the input operation of the target user in the search box of the music application, obtain the content input by the user as the search term, and after obtaining the search term, obtain multiple playlists related to the search term through a search engine or a query database, and through a similarity calculation algorithm. At the same time, the user portrait of the target user can be obtained, so as to recommend a playlist that meets the needs of the target user in the future. Among them, the user portrait can be pre-built, and the target user's listening habits can be tracked through the basic information when the target user registers or logs in, and the user portrait can be built.

[0072] Step S120, for each target playlist, respectively obtain the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under the preset dimension.

[0073] The attribute feature is a specific description of the target playlist's own attributes, which can reflect the characteristics and quality of the target playlist. For example, the attribute feature may include the release time, number of songs, number of comments, and playback volume of the songs in the target playlist.

[0074] Among them, user portrait is a comprehensive description of the user's personal preferences, behaviors and characteristics. For example, the user portrait may include the user's basic information (such as age, gender, region, etc.), music preferences (such as the user's favorite music style, artist, frequently listened songs, etc.), listening habits (such as the frequency and duration of listening to songs at different time periods, etc.), and interactive information (such as the user's interaction, sharing and evaluation of music in the music application).

[0075] Among them, the matching degree between the target playlist and the user portrait includes the matching degree between the songs, artists, genres and other information in the target playlist and the user portrait, which reflects the overlap between the target playlist and the current target user's preferences.

[0076] The conversion rate indicates the ratio of the number of successful conversions to the total number of conversions, which can be click-through rate, play rate, collection rate, sharing rate, etc., and can represent the popularity of the playlist. Therefore, the historical conversion rate of the target playlist under the preset dimensions can represent the popularity of the target playlist under different dimensions.

[0077] In one embodiment, the preset dimensions may include at least one of the music market dimension, the search term dimension, and the target user group dimension to which the target user belongs. Among them, the historical conversion rate under the search term dimension, in addition to characterizing the popularity under this dimension, can also indirectly reflect the matching degree between the target playlist and the search term. Generally speaking, the higher the matching degree, the higher the historical conversion rate will be, and the corresponding popularity will be higher. Similarly, the historical conversion rate under the target user group dimension, in addition to characterizing the popularity under this dimension, can also indirectly reflect the preference of the target playlist in the target user group. The higher the preference, the higher the historical conversion rate will be, and the corresponding popularity will be higher. The historical conversion rate of the target playlist in different dimensions can achieve a more accurate characterization of the target playlist, thereby improving the accuracy of the subsequent target playlist sorting results.

[0078] In a specific implementation, for each target playlist, the information of the songs in the target playlist can be obtained to obtain the attribute characteristics, and the matching degree between the target playlist and the user portrait can be calculated by a similarity calculation method (such as cosine similarity), and the historical conversion rate can be calculated based on the historical interaction data of the target playlist under the preset dimension. This allows the target playlists to be sorted according to their attribute characteristics, matching degree, and historical conversion rate.

[0079] Step S130, input the attribute characteristics, matching degree and historical conversion rate of each target playlist into the pre-trained playlist sorting model to obtain the recommendation value of each target playlist.

[0080] In a specific implementation, a neural network model can be pre-constructed as a playlist sorting model. Then, the historical search terms of the sample user, the historical playlist search results under the historical search terms, and the interaction information of the historical user on the historical playlist search results are obtained, and the label of each historical playlist is marked as a sample label, and this information is combined into sample data. And the sample attribute characteristics of each historical playlist, the sample matching degree with the user portrait of the sample user, and the historical conversion rate under the above preset dimensions are obtained, and then the sample attribute characteristics, sample matching degree and historical conversion index value of the historical playlist are used as input variables, and the sample label is used as the supervision information to train the constructed playlist sorting model to obtain a pre-trained playlist sorting model. Furthermore, the attribute characteristics, matching degree and historical conversion rate of each target playlist can be input into the pre-trained playlist sorting model in turn, and the playlist sorting model calculates the recommendation value of each target playlist, which can be understood as a score.

[0081] Step S140, sorting multiple target playlists according to the recommendation value of each target playlist, and obtaining a playlist sorting result generated for the search term.

[0082] In a specific implementation, the recommendation value can represent the degree of recommendation for the target playlist, and the larger the recommendation value, the higher the recommendation degree. Therefore, after obtaining the recommendation value of each target playlist, the target playlists can be sorted in descending order according to the recommendation value to obtain the playlist sorting result, which is displayed to the target user as the search display result of the search term input by the target user.

[0083] In the above-mentioned method for sorting target playlists, the search terms input by the target user and multiple target playlists associated with the search terms are obtained, and the attribute characteristics of each target playlist, the degree of matching with the user portrait of the target user, and the historical conversion rate under the preset dimensions are respectively obtained. Subsequently, the playlist recommendation model sorts the playlists based on the attribute characteristics, matching degree, and historical conversion rate of each target playlist. This method combines the degree of matching between the target playlist and the user portrait, as well as the attribute characteristics of the target playlist to sort the playlist, and can recommend playlists that meet the user's interests, achieving personalized and accurate playlist recommendations. At the same time, combined with the historical conversion rate of the playlist under the preset dimension to sort the playlist, it can recommend playlists with higher historical conversion rates. Through this multi-dimensional feature-based playlist sorting, the accuracy of the playlists displayed to the user can be improved.

[0084] In an exemplary embodiment, the historical conversion rate of each target playlist under the preset dimension is obtained in the following manner: obtaining the historical interaction data of the target playlist under the preset dimension; and obtaining the historical conversion rate of the target playlist under the preset dimension based on the historical interaction data.

[0085] Among them, the historical interaction data of a target playlist may include the exposure of the target playlist and the user's interaction with the target playlist (such as clicks, favorites, shares, etc.).

[0086] Specifically, the historical conversion rate may include click-through rate, play rate, collection rate, sharing rate, etc., wherein click-through rate = number of clicks / number of exposures, play rate = number of plays / number of exposures, collection rate = number of collections / number of plays, and sharing rate = number of shares / number of plays. According to these relationship formulas, for the calculation of the historical conversion rate of any target playlist in any preset dimension, after obtaining the historical interaction data of the target playlist under the preset dimension, the historical interaction data can be statistically analyzed to obtain conversion parameters, and the historical conversion rate can be obtained based on the conversion parameters. For example, the number of clicks and exposures are counted as conversion parameters to calculate the click-through rate; the number of plays and exposures are counted as conversion parameters to calculate the play rate.

[0087] In this embodiment, the historical interaction data of the target playlist under the preset dimension is obtained; based on the historical interaction data, the historical conversion rate of the target playlist under the preset dimension is statistically obtained to more accurately characterize the target playlist, thereby improving the accuracy of subsequent target playlist sorting results.

[0088] It is understandable that the historical conversion rate under the preset dimension is a conversion feature obtained based on historical interaction data, but the confidence of some historical conversion rates directly calculated through historical interaction data may be affected by the size of the data. For example, although the play rate of 5 times played on the basis of 1 exposure and 50 times played on the basis of 10 exposures is 500%, the confidence of the latter should actually be higher. Therefore, it is necessary to correct the historical conversion rate calculated directly. However, commonly used correction methods such as Wilson smoothing and Bayesian smoothing require the modeling object to satisfy the binomial distribution, that is, the numerator cannot be greater than the denominator, and many conversion features in the playlist scenario do not meet this requirement. For example, a playlist is exposed once, but can be played 5 times. When calculating the conversion rate from exposure to play, the number of plays as the numerator is 5, which is greater than the number of exposures as the denominator. 1. Therefore, for conversion rates that meet the binomial distribution (such as collection rate, sharing rate), Wilson smoothing, Bayesian smoothing and other smoothing methods can be used for correction. For conversion rates that do not meet the binomial distribution (such as click-through rate, play rate), an improved correction method needs to be adopted. The two situations are described below.

[0089] Calculation of historical conversion rate for which the number of successful conversions conforms to the binomial distribution: In an exemplary embodiment, based on the historical interaction data, the historical conversion rate of the target playlist under a preset dimension is obtained by statistics, including: determining whether the number of successful conversions corresponding to the historical conversion rate conforms to the binomial distribution; if it conforms to the binomial distribution, the historical interaction data of the target playlist under the preset dimension is statistically analyzed to obtain conversion parameters, and the conversion parameters are processed by a smoothing estimation method to obtain the historical conversion rate of the target playlist under the preset dimension.

[0090] Specifically, the binomial distribution represents the discrete probability distribution of the number of successes in n independent success / failure tests. If the number of successful conversions corresponding to the historical conversion rate conforms to the binomial distribution, it can be corrected according to the confidence level using commonly used smoothing estimation methods such as Wilson smoothing and Bayesian smoothing. If it does not conform to the binomial distribution, it cannot be directly processed using methods such as Wilson smoothing and Bayesian smoothing.

[0091] Specifically, if it conforms to the binomial distribution, the number of successful conversions and the total number of conversions can be obtained through historical interaction data statistics as conversion parameters, and then the historical conversion rate can be calculated in combination with the smoothing estimation method. For example, taking Wilson smoothing as an example, after obtaining the conversion parameters, select the confidence level, calculate the standard normal score corresponding to the confidence level, and finally use the Wilson interval calculation formula to calculate the corrected historical conversion rate.

[0092] This embodiment takes into account the impact of the amount of data on the confidence of the historical conversion rate, and therefore corrects the historical conversion rate. At the same time, it distinguishes between correction methods that conform to the binomial distribution and those that do not conform to the binomial distribution, which can improve the accuracy and reliability of the acquired historical conversion rate.

[0093] For the calculation of the historical conversion rate of the number of successful conversions that conforms to the binomial distribution: the first processing method is to refer to the Bayesian smoothing method and use the historical values ​​of the large category when introducing the prior numerator and denominator. Correspondingly, in an exemplary embodiment, the above method also includes: if it does not conform to the binomial distribution, then the average conversion parameters of all historical playlists are counted according to the total historical interaction data of all historical playlists under the preset dimension, and the conversion parameters of the target playlist are counted according to the historical interaction data of the target playlist under the preset dimension; the historical conversion rate of the target playlist under the preset dimension is calculated through the average conversion parameters of all target playlists and the conversion parameters of the target playlist.

[0094] In the specific implementation, for the calculation of the historical conversion rate when the number of successful conversions does not conform to the binomial distribution, the total historical interaction data of all playlists under the preset dimension can be obtained first, and the average conversion parameters of all target playlists can be counted based on the total historical interaction data, and the conversion parameters of the target playlist can be counted based on the historical interaction data of the target playlist under the preset dimension. Further, the historical conversion rate of the target playlist under the preset dimension is calculated by the average conversion parameters of all target playlists and the conversion parameters of the target playlist. Among them, the conversion parameters include the number and total number of successful conversions, that is, the parameters on the numerator and denominator. The historical conversion rate is calculated by the average conversion parameters of all target playlists and the conversion parameters of the target playlist, which is obtained by dividing the sum of the numerator in the average conversion parameter and the numerator in the conversion parameter of the target playlist by the sum of the denominator in the average conversion parameter and the denominator in the conversion parameter of the target playlist.

[0095] Taking the play rate, i.e. the conversion rate from exposure to play (CVR), as an example, calculate the historical conversion rate under a certain search term dimension and obtain the average exposure times base of all target playlists under the search term. avg and average number of plays count avg , as the average conversion parameter, and obtain the exposure count base of a single playlist real and play count real , as the transformation parameters of the playlist, these parameters are substituted into the following formula and corrected in the form of Bayesian smoothing.

[0096]

[0097] Among them, C is the adjustment coefficient, which is used to avoid the extreme situation where the base value is too small.

[0098] This embodiment takes into account that in the playlist scenario, due to the small number of exposures of some playlists, the calculation of the conversion rate (such as the conversion rate from exposure to play) may be highly volatile. Therefore, by introducing the historical values ​​of large categories as priors, this volatility can be effectively reduced, and the stability and reliability of the conversion index can be improved. This can achieve smoothing of conversion features that do not satisfy the binomial distribution, and has a wider range of adaptability.

[0099] The second processing method is to convert the number of successful conversions corresponding to the historical conversion rate into a binomial distribution, and then perform smoothing. Correspondingly, in an exemplary embodiment, the above method also includes: if it does not conform to the binomial distribution, then based on the historical interaction data of the target playlist under the preset dimension, the number of successful conversions corresponding to the historical conversion rate is modeled as multiple binomial distribution models; multiple binomial distribution models are smoothed to obtain the historical conversion rate of the target playlist under the preset dimension.

[0100] In the specific implementation, first define the highest number of successful conversions of the target playlist under a single exposure based on the historical interaction data of the target playlist under the preset dimensions; based on the historical interaction data, count the frequency of each conversion under the highest conversion number to obtain multiple binomial distribution models corresponding to the number of the highest conversion number. Further smooth the multiple binomial distribution models to obtain the historical conversion rate of the target playlist under the preset dimensions.

[0101] In an exemplary embodiment, the above-mentioned step of modeling the number of successful conversions corresponding to the historical conversion rate as multiple binomial distribution models based on the historical interaction data of the target playlist under preset dimensions specifically includes: defining the maximum number of successful conversions of the target playlist under a single exposure; counting the number of successful conversions of the target playlist at each exposure based on the historical interaction data; and separately counting the frequency of occurrence of each successful conversion number that is less than the maximum successful number to construct multiple binomial distribution models.

[0102] Specifically, each display of a playlist is a single exposure. For example, if a playlist is displayed 100 times in one day, there will be 100 single exposures. If 50 users click on it in the first display, the number of successful conversions is 50, and if 100 people click on it in the second display, the number of successful conversions is 100. Each exposure has a corresponding number of successful conversions. The maximum number of successful conversions K for a single exposure can be limited by truncation, and then the counts of conversions 1, 2, and up to K times are counted respectively, thereby converting them into modeling of K binomial distributions, that is, obtaining K binomial distribution models. These K binomial distributions can be corrected using the Wilson smoothing method mentioned above. After obtaining the K binomial distribution models, they can be used directly as features, as shown in formula (2), or they can be weighted summed to obtain the corrected conversion results, as shown in formula (3).

[0103] CVR k = Wilson(count, base tuned ) (2)

[0104] CVR weighted =∑k×CVR k (3)

[0105] Among them, CVR k Indicates the transformation result with count k times; Wilson indicates the Wilson smoothing method; count and base tuned Indicates the conversion parameter with count k, corresponding to the parameters on the numerator and denominator respectively; CVR weighted It represents the conversion result obtained by weighted summing of K conversion results, and k in formula (3) represents the weight corresponding to the count of k times.

[0106] It should be noted that for the above calculation of the historical conversion rate that does not conform to the binomial distribution, one of the first processing method and the second processing method can be selected for use, or both can be used, such as adding them together in a weighted form.

[0107] In the above embodiment, for the calculation of historical conversion rates that do not satisfy the binomial distribution, the number of successful conversions is first modeled as a binomial distribution so that it can better conform to statistical principles, and then smoothed. This can effectively reduce random fluctuations caused by small sample sizes or uneven data distribution, thereby improving the accuracy and reliability of conversion rate estimation, which is conducive to optimizing business processes.

[0108] In an exemplary embodiment, if Figure 2 As shown in the figure, the playlist sorting model is trained in the following way:

[0109] Step S210, obtaining sample data; the sample data includes historical search terms of the sample user, historical playlists associated with the historical search terms, and interactive data of the sample user on the search results of the historical playlists; the historical playlists have corresponding sample tags;

[0110] Step S220, obtaining sample attribute characteristics of historical playlists, sample matching degree with user portraits of sample users, and historical conversion rates under preset dimensions;

[0111] Step S230, using the sample attribute characteristics, sample matching degree and historical conversion rate of the historical playlist as input variables and the sample label as supervision information, the initial playlist sorting model is trained to obtain a pre-trained playlist sorting model.

[0112] In the specific implementation, Figure 3 As shown in the figure, the training of the playlist sorting model mainly includes several steps: (1) Collect search logs and associate the reported data from various sources. (2) Construct sample labels, perform positive and negative annotations and sample sampling based on search logs. (3) Calculate the conversion rate of the playlist in the sample, and summarize and calculate feature data from various dimensions such as search terms, music market, user groups, playlists, and matching degree with user portraits. (4) Train the model and deploy it online, continuously and automatically update the features so that the sorting responds in a timely manner. The following is a detailed description of each step.

[0113] (1) Collect search logs. Data can be reported in advance on the music application side, such as the search terms (query) entered by the user in the search box, the sorting results of the playlist resources returned after the user initiates the search, and which playlists the user clicks on or other further operations on the secondary page. Since these data are reported separately in an asynchronous form, for example, the exposure results returned to the client will be reported separately when the search results are returned, and then stored in the exposure data table; the click behavior will be reported separately after the client clicks, and then stored in the click report data table; the playback behavior will be reported after a playback ends or is switched out, and then the playback duration and other information will be stored in the playback report data table. Therefore, when the user initiates a search, the background will generate a sid as an identifier for the current search, and it will run through the entire link, so that these data can be associated. Such as Figure 4 As shown, after these reported data are recovered, association processing can be performed. The ultimate goal is to restore as much as possible the search results and subsequent operation sequence seen by the user each time the user searches, as well as some detailed information about the operation object, that is, the sample user's historical search terms, the historical playlist search results under the historical search terms, and the sample user's interactive information on the historical playlist search results.

[0114] (2) Construct sample labels. After obtaining the associated search logs, it is necessary to annotate and collect positive and negative samples based on the user behavior records of each search. Taking a user's playlist search log as an example, it can be expanded in the order of exposure position, so as to more clearly know which playlists the user clicked in the current results, and how many songs were played on the secondary pages of these playlists, how long they were played, how many songs were played, how many songs were played, and how many songs were played. Then, the positive and negative samples are annotated. Different from other non-music search scenarios, the ultimate goal of playlist ranking optimization is to improve the user's satisfaction experience in search results. The specific quantitative indicators will reflect the user's complete playback behavior and some information including collections and comments in each search. The user's click on the search results page does not necessarily represent true positive feedback. Only when "effective playback" and satisfactory clicks are performed on the playlist will it be considered positive feedback. Therefore, the annotation is divided into positive sample annotation and negative sample annotation. Furthermore, both positive and negative sample annotations can be divided into multiple levels, as shown in Table 1 below. The following positive sample annotation principles can be formulated: 1. If there is one completion behavior or a satisfactory click on the song granularity and the non-completed playback does not exceed the specified number of times, it is considered to be the first level of satisfaction. Here, some users have listened to the song multiple times, which means that the songs in the playlist cannot meet the user's needs; 2. If there are three completions and no multiple attempts to play, it is considered to be the second level of satisfaction; 3. If there are more than 5 completions or a satisfactory behavior at the granularity of the playlist, it is considered to be the third level of satisfaction. Since the user has entered an immersed playback state at this time, it is considered to be the most in line with the user's needs. For negative samples, in addition to the playlists that are generally exposed but not operated, clicks without playback and clicks and multiple incomplete playbacks will be regarded as negative samples. The labels of the former two will use the 0 label, while the latter will use the -1 label, indicating a negative sample with a worse feedback level.

[0115] (3) Calculate the conversion rate of the playlists in the sample. Similar to the above step S120 of obtaining multiple playlists associated with the search term, and obtaining the attribute characteristics of each playlist, the matching degree with the user portrait, and the conversion rate under the preset dimension, when processing the sample data, it is also necessary to calculate these conversion rates and correct the conversion rates. The relevant calculation and correction methods are similar to those in the above embodiment and will not be repeated here.

[0116] Table 1. Song list sample annotation method

[0117] Sample Type Feedback conditions Sample Label Strong Negative Samples Tried to listen many times but never finished playing -1 Weak Negative Samples No Action 0 First-level positive sample Click when the song is finished or you are satisfied with it. 1 Secondary positive samples Three times of playing or satisfactory clicks on the song 2 Three-level positive samples Five complete plays or one satisfactory click on the playlist 3

[0118] (4) Model training and deployment. After the sample annotation and feature calculation are completed, the complete training data can be associated and then the model can be trained. The model can be selected from traditional machine learning or deep model algorithms, such as sorting models based on point-wise (point pair sorting method, mainly considering the relationship between a single query and a document, converting the problem into a multi-classification or regression problem), pair-wise (document pair sorting method, by comparing each pair of documents (doc pair) to determine their relative order), list-wise (document list sorting method, considering the entire document list as a whole, and directly taking document sorting as the optimization target). During training, the sample attribute characteristics, sample matching degree and historical conversion rate of the historical playlist are used as input variables, and the sample label is used as supervision information to train the initial playlist sorting model to obtain a pre-trained playlist sorting model. For example, taking the XGBOOST (an optimized distributed gradient boosting library) sorting model as an example, in the prediction stage, the attribute characteristics, matching degree and conversion rate of the current playlist to be predicted are input into the decision tree for weighting to obtain a predicted recommendation value, which represents the predicted sorting relationship. Through the gradient boosting method of the tree model (in each iteration, the model calculates the gradient between the current prediction value and the true label, generates a residual (wrong prediction) for each sample, so that the training of subsequent trees can focus on correcting these errors. This residual can be approximately represented by a negative gradient), these trees are generated by continuously selecting the feature splitting point with the maximum benefit (the maximum benefit here is the approximate solution, and the goal is to find the direction that can make the loss function drop the fastest. The loss function is the error between the aforementioned residual and the result of the tree to be generated in the current round, plus some regular terms) to generate branches and finally generate a decision tree, and then multiple rounds of iterative corrections are performed to generate a batch of decision trees to complete the training.

[0119] After training the playlist sorting model offline, the model can be deployed online. In order to ensure that the sorting effect can satisfy users in a timely manner, the feature data stream also needs to be continuously and stably updated. For the training phase, the synchronous association can be pulled separately. There are many sources of data available online, such as user portrait data is updated in real time, playlist attribute data is updated irregularly, and playlist history data is regularly counted. Therefore, this application designs a set of asynchronous association feature methods that integrate various sources, such as Figure 5As shown in the figure, user portrait data is obtained through real-time interface requests, playlist attribute data has low real-time performance and is pulled and stored on the server through scheduled tasks, while playlist history data is read, written and updated through a massive distributed storage system (each business has a unique identifier (tid) and a routing table, which describes the location information of the identifier's key value stored in the Cache). The latter two only store count data, and the calculation logic is placed in the server, which can reduce the pressure of offline calculations and complete feature updates faster.

[0120] The song list sorting model training method provided in this embodiment is applicable to various models in the industry. Based on the feature construction and sample annotation method proposed in this solution, high-quality training data can be obtained, and high-quality sorting models can be obtained using different traditional machine learning or deep learning models. By designing positive and negative feedback sample annotation and construction methods, it can better reflect the real preferences of users in music search scenarios; by designing song list feature calculation and smoothing processing methods, the quality of song lists and user consumption intentions can be better portrayed; by designing a set of online real-time automatic updating model sorting solutions, the sorting effect can be improved while reducing labor costs.

[0121] In an exemplary embodiment, step S210, obtaining sample data, includes: obtaining the average number of playlists that can be displayed in the playlist display page, and constructing a playlist sampling window according to the average number of playlists; determining a playlist sampling window with positive samples, and within the sampling window with positive samples, performing playlist sampling in the form of probability attenuation to obtain negative samples; and combining the positive samples and the negative samples into sample data.

[0122] It is understandable that after labeling the positive and negative samples, the next step of sampling is required. Since the number of exposed playlists reported in the search log each time can be as many as dozens or even hundreds, and there may be only a few or even only one positive sample labeled, not sampling will make the ratio of positive and negative samples too disparate, which is not conducive to model training. In fact, users may not necessarily observe all the exposed playlists reported, so negative samples need to be sampled. Figure 6 As shown in the figure, the fixed window size, i.e., the size of the upper and lower samples of the positive sample, can be determined based on the average number of playlists that can be displayed on a single screen return page in the client. At the same time, a number of playlists are sampled in the sampling window in the form of probability decay, so that the sampled positive and negative samples can be controlled at a certain ratio, and at the same time, it can be ensured as much as possible that the positive and negative samples are actually seen by the user, thereby reflecting the real user preference tendency.

[0123] In this embodiment, by performing probability attenuation sampling of negative samples in the sampling window marked with positive samples, the ratio of positive and negative samples can be effectively balanced. This helps to improve the model's learning ability on different types of samples and reduce the risk of the model being biased towards positive or negative samples. By obtaining the average number of playlists that the client can display on a single page and using this as the sampling window, sample bias can be reduced to ensure that the sampling process is more in line with the user's actual usage scenario.

[0124] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0125] Based on the same inventive concept, the embodiment of the present application also provides a song list sorting device for implementing the song list sorting method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more song list sorting device embodiments provided below can refer to the limitations on the song list sorting method above, and will not be repeated here.

[0126] In one embodiment, Figure 7 As shown, a song list sorting device is provided, comprising:

[0127] The term acquisition module 710 is used to acquire the search term input by the target user and multiple target playlists associated with the search term;

[0128] The feature acquisition module 720 is used to obtain, for each target playlist, the attribute features of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension; the historical conversion rate represents the ratio of the number of successful conversions to the total number of conversions;

[0129] The recommendation value generation module 730 is used to input the attribute characteristics, matching degree and historical conversion rate of each target playlist into the pre-trained playlist sorting model to obtain the recommendation value of each target playlist;

[0130] The playlist sorting module 740 is used to sort multiple target playlists according to the recommendation value of each target playlist to obtain a playlist sorting result generated for the search term.

[0131] In one of the embodiments, the preset dimensions include at least one of a music market dimension, a search term dimension, and a target user group dimension to which the target user belongs. The historical conversion rate under the music market dimension reflects the popularity of the target playlist in the music market, the historical conversion rate under the search term dimension reflects the matching degree between the target playlist and the search term, and the historical conversion rate under the target user group dimension reflects the preference of the target playlist in the target user group.

[0132] In one of the embodiments, the feature acquisition module 720 is also used to obtain historical interaction data of the target playlist under a preset dimension; based on the historical interaction data, the historical conversion rate of the target playlist under the preset dimension is statistically obtained.

[0133] In one of the embodiments, the feature acquisition module 720 is also used to determine whether the number of successful conversions corresponding to the historical conversion rate conforms to the binomial distribution; if it conforms to the binomial distribution, the historical interaction data of the target playlist under the preset dimension is statistically analyzed to obtain the conversion parameters; the conversion parameters include the number of successful conversions and the total number of conversions; the conversion parameters are processed by a smoothing estimation method to obtain the historical conversion rate of the target playlist under the preset dimension.

[0134] In one of the embodiments, the feature acquisition module 720 is also used to count the average conversion parameters of all historical playlists based on the total historical interaction data of all historical playlists under the preset dimension if it does not conform to the binomial distribution, and to count the conversion parameters of the target playlist based on the historical interaction data of the target playlist under the preset dimension; calculate the historical conversion rate of the target playlist under the preset dimension through the average conversion parameters of all historical playlists and the conversion parameters of the target playlist.

[0135] In one of the embodiments, the feature acquisition module 720 is also used to model the number of successful conversions corresponding to the historical conversion rate into multiple binomial distribution models based on the historical interaction data of the target playlist under the preset dimension if it does not conform to the binomial distribution; and smooth the multiple binomial distribution models to obtain the historical conversion rate of the target playlist under the preset dimension.

[0136] In one of the embodiments, the feature acquisition module 720 is also used to define the maximum number of successful conversions of the target playlist under a single exposure; count the number of successful conversions of the target playlist at each exposure based on historical interaction data; and count the frequency of occurrence of each successful conversion number that is less than the maximum successful number to construct multiple binomial distribution models.

[0137] In one of the embodiments, the device also includes a model training module for obtaining sample data; the sample data includes: historical search terms of sample users, historical playlists associated with the historical search terms, and interaction data of sample users on historical playlists; historical playlists have corresponding sample labels; sample attribute characteristics of historical playlists, sample matching degrees with user portraits of sample users, and historical conversion rates under preset dimensions are obtained; the sample attribute characteristics, sample matching degrees, and historical conversion rates of historical playlists are used as input variables, and the sample labels are used as supervision information to train the initial playlist sorting model to obtain a pre-trained playlist sorting model.

[0138] In one of the embodiments, the model training module is also used to obtain the average number of playlists that can be displayed in the playlist display page, and construct a playlist sampling window based on the average number of playlists; determine the playlist sampling window with positive samples, and within the playlist sampling window with positive samples, perform playlist sampling in the form of probability attenuation to obtain negative samples; and combine the positive samples and negative samples into sample data.

[0139] Each module in the above-mentioned song list sorting device can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0140] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a song list sorting method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0141] Those skilled in the art will understand that Figure 8The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0142] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0143] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0144] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0145] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0146] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0147] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0148] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A song list sorting method, characterized in that: The method comprises: Acquire a search term input by a target user and a plurality of target playlists associated with the search term; For each of the target playlists, respectively obtain the attribute characteristics of the target playlist, the matching degree between the target playlist and the user portrait of the target user, and the historical conversion rate of the target playlist under a preset dimension; Inputting the attribute features, the matching degree and the historical conversion rate of each target playlist into a pre-trained playlist sorting model respectively to obtain a recommendation value for each target playlist; The multiple target playlists are sorted according to the recommendation value of each of the target playlists to obtain a playlist sorting result generated for the search term.

2. The method according to claim 1, characterized in that The preset dimensions include at least one of a music market dimension, a search term dimension, and a target user group dimension to which the target user belongs. The historical conversion rate under the music market dimension reflects the popularity of the target playlist in the music market, the historical conversion rate under the search term dimension reflects the matching degree between the target playlist and the search term, and the historical conversion rate under the target user group dimension reflects the preference of the target playlist in the target user group.

3. The method according to claim 1, characterized in that Obtaining the historical conversion rate of the target playlist under the preset dimension includes: Obtain historical interaction data of the target playlist under the preset dimension; According to the historical interaction data, the historical conversion rate of the target playlist under the preset dimension is obtained by statistics.

4. The method according to claim 3, characterized in that The step of obtaining the historical conversion rate of the target playlist under the preset dimension based on the historical interaction data includes: Determine whether the number of successful conversions corresponding to the historical conversion rate conforms to a binomial distribution; If it conforms to the binomial distribution, the historical interaction data of the target playlist under the preset dimension is statistically analyzed to obtain conversion parameters; the conversion parameters include the number of successful conversions and the total number of conversions; The conversion parameters are processed by a smoothing estimation method to obtain the historical conversion rate of the target playlist under the preset dimension.

5. The method according to claim 4, characterized in that The method further comprises: If it does not conform to the binomial distribution, then the average conversion parameter of all historical playlists is calculated based on the total historical interaction data of all historical playlists under the preset dimension, and the conversion parameter of the target playlist is calculated based on the historical interaction data of the target playlist under the preset dimension; The historical conversion rate of the target playlist under the preset dimension is calculated by using the average conversion parameter of all the historical playlists and the conversion parameter of the target playlist.

6. The method according to claim 4, characterized in that The method further comprises: If it does not conform to the binomial distribution, the number of successful conversions corresponding to the historical conversion rate is modeled as multiple binomial distribution models according to the historical interaction data of the target playlist under the preset dimension; The multiple binomial distribution models are smoothed to obtain the historical conversion rate of the target playlist under the preset dimension.

7. The method according to claim 6, characterized in that According to the historical interaction data of the target playlist under the preset dimension, the number of successful conversions corresponding to the historical conversion rate is modeled as multiple binomial distribution models, including: Define the maximum number of successful conversions of the target playlist in a single exposure; According to the historical interaction data, the number of successful conversions of the target playlist at each exposure is counted; The occurrence frequency of each conversion success number that is less than the highest success number is counted respectively to construct the multiple binomial distribution models.

8. The method according to claim 1, characterized in that The playlist sorting model is trained in the following way: Acquire sample data; the sample data includes: historical search terms of a sample user, historical playlists associated with the historical search terms, and interaction data of the sample user on the historical playlists; the historical playlists have corresponding sample tags; Obtaining sample attribute characteristics of the historical playlist, sample matching degree with the user portrait of the sample user, and historical conversion rate under preset dimensions; The sample attribute characteristics of the historical playlist, the sample matching degree and the historical conversion rate are used as input variables, and the sample labels are used as supervision information to train the initial playlist sorting model to obtain the pre-trained playlist sorting model.

9. The method according to claim 8, characterized in that The obtaining of sample data comprises: Obtain the average number of playlists that can be displayed on the playlist display page, and construct a playlist sampling window according to the average number of playlists; Determine a song list sampling window with positive samples, and perform song list sampling in the form of probability attenuation within the song list sampling window with positive samples to obtain negative samples; The positive samples and the negative samples are combined into the sample data.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the playlist sorting method described in any one of claims 1 to 9 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the playlist sorting method described in any one of claims 1 to 9 are implemented.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the playlist sorting method described in any one of claims 1 to 9 are implemented.

Images