Resource sorting method, device, computing device and storage medium

Abstract

The embodiment of the present disclosure provides a resource sorting method, apparatus, computing device and storage medium, the method comprising: obtaining multiple resource modules according to a resource request, wherein each resource module is heterogeneous and each resource module includes at least one resource; obtaining online feature information, the online feature information at least including the first feature information of the resource module and the second feature information of the resource request; inputting the online feature information into a resource sorting network to obtain the sorting results of the multiple resource modules; and performing recommendation display of the multiple resource modules according to the sorting results of the multiple resource modules output by the resource sorting network. By merging the features of multiple resource modules and the features of resource requests into online feature information, and using the resource sorting network to make a unified prediction of each resource module, the influence between each resource module can be captured, and the accuracy of the sorting results is improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of data processing technology. More specifically, the embodiments of the present disclosure relate to a resource sorting method, apparatus, computing device and storage medium. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims. No description herein is admitted to be prior art by inclusion in this section.

[0003] In the music search scenario, a lot of resources are usually involved, including songs, playlists, videos, podcasts, sounds, albums, artists, search terms, users, etc. The recalled resources are displayed in the form of modules on the search comprehensive page. The form and display order of these modules are crucial to satisfying the user's search intent, and are also the challenges of multi-resource mixing technology in the current music scenario.

[0004] In the related technology, a scheme of independently scoring modules by training machine learning models can be adopted to sort the modules. One way is to use multiple modules to share a model and score them separately. This method treats multiple heterogeneous modules as homogeneous modules and treats models containing different resources uniformly. It is inconsistent with the actual business and it is difficult to capture the individual characteristics of each module. Another way is to use each module to train its own model separately and then score them separately. In this way, the scoring standards of each model may be inconsistent, the final score is less comparable, and it is difficult to adjust the scores of each module later.

[0005] Public Content

[0006] In view of this, the embodiments of the present disclosure at least provide a resource recommendation method, apparatus, computing device and storage medium to perform unified real-time personalized sorting of resource modules in a scenario of searching for multiple heterogeneous resources.

[0007] In a first aspect of the embodiments of the present disclosure, a resource recommendation method is provided, which is applied to an online system and includes:

[0008] Acquire multiple resource modules according to the resource request, wherein the resource modules are heterogeneous and each resource module includes at least one resource;

[0009] Acquire online characteristic information, where the online characteristic information at least includes first characteristic information of the resource module and second characteristic information of the resource request;

[0010] Inputting the online feature information into a resource sorting network to obtain sorting results of the multiple resource modules;

[0011] According to the sorting results of the multiple resource modules output by the resource sorting network, the multiple resource modules are recommended and displayed.

[0012] In a second aspect of the embodiments of the present disclosure, a method for training a resource sorting network is provided, which is applied to an offline system and includes:

[0013] Obtain client logs and server logs of online systems;

[0014] Obtaining a labeled data set and offline information according to the client log, the server log, and the first resource request information stored offline, wherein the resource module in the labeled data set is labeled with a consumption real value indicating whether the resource module is consumed after being exposed;

[0015] Acquire offline feature information according to the offline information;

[0016] Generate a sample data set according to the offline feature information and the annotated data set;

[0017] The resource sorting network is trained offline according to the sample data set.

[0018] In a third aspect of the embodiments of the present disclosure, a resource recommendation device is provided, which is applied to an online system and includes:

[0019] A recall unit, configured to obtain a plurality of resource modules according to a resource request, wherein the resource modules are heterogeneous and each of the resource modules includes at least one resource;

[0020] A feature extraction unit, which obtains online feature information, wherein the online feature information at least includes first feature information of the resource module and second feature information of the resource request;

[0021] A sorting unit, used for inputting the online feature information into a resource sorting network to obtain sorting results of the multiple resource modules;

[0022] A recommendation unit is used to recommend and display the multiple resource modules according to the sorting results of the multiple resource modules output by the resource sorting network.

[0023] In a fourth aspect of the embodiments of the present disclosure, a training device for a resource sorting network is provided, which is applied to an offline system and includes:

[0024] Log acquisition unit, used to obtain client logs and server logs of online systems;

[0025] A data acquisition unit, configured to obtain a labeled data set and offline information according to the client log, the server log, and the first resource request information stored offline, wherein the offline information at least includes resource module information and second resource request information, and the resource module in the labeled data set is labeled with a consumption real value indicating whether the resource module is consumed after being exposed;

[0026] A feature acquisition unit, configured to acquire offline feature information according to the offline information, wherein the offline feature information at least includes fourth feature information of the resource module and fifth feature information of the resource request;

[0027] A sample generating unit, configured to generate a sample data set according to the offline feature information and the annotated data set;

[0028] A training unit is used to perform offline training on the resource sorting network according to the sample data set.

[0029] In a fifth aspect of the embodiments of the present disclosure, a computing device is provided, including: a processor; and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method described in any embodiment of the present disclosure is implemented.

[0030] In a sixth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, comprising: a computer program, wherein when the computer program is executed by a processor, the method described in any embodiment of the present disclosure is implemented.

[0031] According to the resource recommendation method, apparatus, computing device and storage medium of the embodiment of the present disclosure, multiple heterogeneous resource modules are obtained according to resource requests; online feature information is obtained, and the online feature information at least includes first feature information of the resource module and second feature information of the resource request; and the online feature information is input into a resource sorting network to obtain sorting results of the multiple resource modules; based on the sorting results of the multiple resource modules output by the resource sorting network, the multiple resource modules are recommended and displayed, the features of the multiple resource modules and the features of the resource request are merged into online feature information, and a resource sorting network is used to predict the multiple resource modules at the same time, the evaluation criteria for each resource module are unified, and the features of each resource module can be captured, thereby improving the accuracy of the sorting results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0033] Figure 1 The flowchart of the resource recommendation method according to the embodiment of the present disclosure is schematically shown;

[0034] Figure 2 A schematic diagram of a network architecture of a first resource sorting network according to an embodiment of the present disclosure is shown;

[0035] Figure 3 A schematic diagram of a network architecture of a second resource sorting network according to an embodiment of the present disclosure is shown;

[0036] Figure 4 The flowchart of a method for training a resource sorting network according to an embodiment of the present disclosure is schematically shown;

[0037] Figure 5 The structure diagram of a resource recommendation device according to an embodiment of the present disclosure is schematically shown;

[0038] Figure 6 The structure diagram of a training device for a resource sorting network according to an embodiment of the present disclosure is schematically shown;

[0039] Figure 7 Schematically illustrates a computer-readable storage medium according to an embodiment of the present disclosure;

[0040] Figure 8 The schematic diagram shows a structure of a computing device according to an embodiment of the present disclosure.

[0041] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0042] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0043] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0044] According to the embodiments of the present disclosure, a method, an apparatus, a computing device and a storage medium are proposed. In this document, it should be understood that any number of elements in the drawings is for illustration and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0045] The data involved in the present disclosure may be data authorized by the user or fully authorized by all parties.

[0046] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure.

[0047] The disclosed embodiment proposes a resource recommendation solution, the purpose of which is to solve the problem of unified real-time personalized sorting of resource modules in the scenario of searching for multiple heterogeneous resources. The resource recommendation solution can be divided into two parts: an online system and an offline system. The offline system is used for data construction, feature extraction, and sorting model training, and the online system is used for scoring and sorting each resource module online.

[0048] Figure 1 The process of a resource recommendation method according to an embodiment of the present disclosure is schematically shown. The resource recommendation method can be executed by a terminal device, which can be a fixed terminal or a mobile terminal, such as a mobile phone, a tablet computer, a game console, a desktop computer, an advertising machine, an all-in-one machine, a vehicle-mounted terminal, etc. The method can also be implemented by a processor calling a computer-readable instruction stored in a memory. Figure 1 As shown, the method may include the following processing:

[0049] In step 101, multiple resource modules are obtained according to a resource request.

[0050] The resource modules are heterogeneous, and each resource module includes at least one resource.

[0051] After the user initiates a resource request, after understanding the intent of the resource request, resources can be recalled to obtain multiple types of resources, and each type of resource can be displayed in the form of a module. Since the resources in different resource modules belong to different types, the modules are heterogeneous.

[0052] Taking the music search scenario as an example, the recalled resource modules may include single resources, playlist resources, video resources, podcast resources, single recommendation resources, album resources, artist resources, sound resources, user resources, mixed recommendation resources, etc.

[0053] In step 102, online characteristic information is obtained, where the online characteristic information at least includes first characteristic information of the resource module and second characteristic information of the resource request.

[0054] The first characteristic information is a separate characteristic of each resource module, and is organized according to the characteristics of the resource module itself. The first characteristic information may include at least one of the following: natural language processing data of the resource module; historical consumption information and / or real-time consumption information of the resource module; text matching information between the resource request and the resource module; a portrait of the resource module and a portrait of each resource in the resource module. The natural language processing data of the resource module includes the ID characteristic code of the resource module, etc.

[0055] The second characteristic information is the characteristic of the resource request currently being made. The second characteristic information may include at least one of the following: the intent information of the resource request; the natural language processing data of the resource request; the historical consumption information and / or real-time consumption information of each resource module under the resource request. Among them, the intent information of the resource request is the result of intent analysis of the search term in the current resource request, including word segmentation, the hit resource name (for example, song name, album name, playlist name, etc.), the static portrait information of the resource request itself, etc.; the historical consumption information and / or real-time consumption information of each resource module under the resource request is the consumption status of various resource modules under the resource request obtained based on logs and historical log statistics.

[0056] In step 103, the online feature information is input into a resource sorting network to obtain sorting results of the multiple resource modules.

[0057] The resource sorting network is obtained by offline training of an offline system. The resource sorting network scores each resource module according to the first characteristic information of each resource module and the second characteristic information of the resource request, and sorts multiple resource modules according to the score to obtain a sorting result.

[0058] In step 104, the plurality of resource modules are recommended and displayed according to the ranking results of the plurality of resource modules output by the resource ranking network.

[0059] In one example, the first feature information of all resource modules may be organized according to the input format of the resource sorting network and concatenated according to a fixed order of resource modules (so as to be consistent with the offline training process).

[0060] For example, the N modules that need to be sorted are {M1, M2, M3, ..., M N}, the first feature information obtained by splicing in a fixed order is [Feat1, Feat2, Feat3, ..., FeatN], and the scores of multiple modules are predicted by the resource sorting network and recorded as {S1, S2, S3, ..., SN}. Finally, the modules are sorted from high to low according to the scores to obtain the sorting results of multiple resource modules.

[0061] In the disclosed embodiment, a plurality of heterogeneous resource modules are obtained according to a resource request; online feature information is obtained, the online feature information comprising at least first feature information of the resource module and second feature information of the resource request; and the online feature information is input into a resource sorting network to obtain sorting results of the plurality of resource modules; based on the sorting results of the plurality of resource modules output by the resource sorting network, the plurality of resource modules are recommended and displayed, the features of the plurality of resource modules and the features of the resource request are merged into online feature information, and a resource sorting network is used to simultaneously predict the plurality of resource modules, the evaluation criteria for each resource module are unified, and the features of each resource module can be captured, thereby improving the accuracy of the sorting results.

[0062] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0063] In some implementations, before inputting the online feature information into a resource ranking network to obtain ranking results of the plurality of resource modules, the method further includes determining whether the search term corresponding to the resource request is a long-tail term.

[0064] In one example, the long-tail word discriminator may use a logistic regression model. The formula used by the logistic regression model is as follows:

[0065] QScore = sigmoid(β+w T v) (1)

[0066]

[0067] Among them, QScore is the score of the search term corresponding to the resource request, β and w are model parameters of the logistic regression model, v represents the characteristics of the resource request, including: real-time search statistics of resource requests (search volume, resource consumption), historical search statistics of resource requests, number of resources recalled by resource requests, etc.; η is a pre-set hyperparameter, which is usually set according to business performance; p represents the judgment result, p=1 indicates that the search term corresponding to the resource request is a long-tail word (long-tail traffic word), and p=0 indicates that the search term corresponding to the resource request is not a long-tail word (non-long-tail traffic word).

[0068] When the search term is not a long-tail term, the online feature information is input into a first resource sorting network to obtain a sorting result of the multiple resource modules; when the search term is a long-tail term, the online feature information is input into a second resource sorting network to obtain a sorting result of the multiple resource modules.

[0069] In the case where the search term is not a long-tail term, the third characteristic information of the user requesting resource recommendation is also obtained, and the third characteristic information includes at least one of the following: the user's historical preference information for the resource module; the user's historical consumption information and / or real-time consumption information for the resource module; the user's search history information. Among them, the user's historical preference information for the resource module mainly reflects the user's medium- and long-term resource type preference. By mining the user's historical logs, the user's preference score for each resource (including songs, playlists, videos, playlists, albums, artists, sounds, cloud circles, users, keywords, etc.) is calculated. The format is, for example, {type1: score1, type2: score2, ..., typeN: ScoreN}, where typei and scorei represent the user's preference for the i-th resource; real-time consumption information mainly reflects the user's short-term real-time resource type preference, which is obtained based on online real-time logs; search history information mainly reflects the user's click, play, favorite and other behaviors in the search scenario, and assists in the sorting of resource modules under this resource request through the user's existing search behavior.

[0070] By adding user-side features to online feature information, the intent of the current resource request and user personalized data can be combined to provide resource module recommendation results that combine resource request intent and user personalization.

[0071] Figure 2 A schematic diagram of the network architecture of a first resource sorting network proposed in at least one embodiment of the present disclosure is shown below. Figure 2 The method for processing the online feature information by the first resource ranking network when the search term is not a long-tail term is described.

[0072] First, the discrete features in each type of feature information in the online feature information are converted into fixed-length vectors, and a Cartesian crossover operation is performed on each fixed-length vector to obtain a first operation result.

[0073] See also Figure 2Each type of feature information in the online feature information includes user historical resource preferences, user real-time resource consumption preferences, user search history in user-side features, intention requests and statistical information on the resource request side, and various types of resource modules on the resource module side, such as single module, song module, video module, user module, etc. Among them, the statistical information on the resource request side includes historical consumption information and / or real-time consumption information of each resource module under the resource request.

[0074] The discrete features in each type of feature information are converted into fixed-length vectors, including using embedding coding to convert the discrete features in each type of feature information into fixed-length vectors; for the sequence of discrete features, using pooling operation to convert them into fixed-length vectors; then, a Cartesian crossover operation is performed on all fixed-length vectors on the user side, the resource request side, and the resource module side to obtain a first operation result. For example, a Cartesian crossover operation can be performed on each fixed-length vector through a Cartesian (FM) crossover layer.

[0075] Next, the continuous features in each type of feature information in the online feature information are spliced ​​to obtain a splicing result, and after splicing the splicing results, a full connection operation is performed to obtain a second operation result.

[0076] The continuous features in each type of feature information in the online feature information are spliced, and a concatenation operation can be used to obtain a concatenation result; then, all the concatenation results on the user side, the resource request side, and the resource module side are concatenated and then a full connection operation is performed to obtain a second operation result.

[0077] In one example, a DNN (Deep Neural Network) network can be constructed by splicing to perform splicing operations and full connection operations. Figure 2 As shown, the splicing structure DNN network may include a full splicing (Contact) layer and two full connection (FC) layers, wherein the first FC layer may be 1024-dimensional (Dim) and the second FC layer may be 256-dimensional. Those skilled in the art should understand that the parameters of the above network structure are only examples and are not limited in this disclosure.

[0078] Afterwards, the first operation result and the second operation result are concatenated, and a full connection operation is performed to obtain the scores of each resource module.

[0079] In one example, the first operation result and the second operation result can be concatenated and fully connected through a fully connected layer network, and finally all resource modules are scored uniformly in the scoring layer through a softmax function to obtain the scores of each resource module {p1, p2, p3, ..., p N},0<p i ≤1.0.

[0080] like Figure 2 As shown, the full link layer network may include a full contact layer and two full connection (FC) layers, wherein the first FC layer may be 256-dimensional and the second FC layer may be 64-dimensional. Those skilled in the art should understand that the parameters of the above network structure are only examples and are not limited in the present disclosure.

[0081] Finally, a ranking result of the multiple resource modules is obtained according to the scores.

[0082] In the disclosed embodiment, in the design of the resource sorting network, the influence between all resource modules is taken into consideration. While obtaining the scores of the resource modules, the cross-relationships between the resource modules under different feature conditions can also be captured, thereby improving the accuracy of the sorting results.

[0083] Figure 3 A schematic diagram of the network architecture of a second resource sorting network proposed in at least one embodiment of the present disclosure is shown below. Figure 3 The method for processing the online feature information by the second resource ranking network when the search term is a long-tail term is described.

[0084] First, the continuous features in the first feature information and the second feature information are concatenated, and then a full connection operation is performed to obtain the score of each resource module.

[0085] Since the user-side feature orientation is not obvious for long-tail words, and the statistics of many statistical value features on the resource request side are not significant, it is easy to cause overfitting. Therefore, in this step, some features on the resource request side (such as Figure 3 The intent information shown) and the continuous features in the features on the resource module side are spliced.

[0086] In one example, the continuous features in the first feature information and the second feature information can be concatenated and fully connected through a fully connected layer network, and finally all resource modules are scored uniformly in the scoring layer through a softmax function to obtain the scores of each resource module {p1, p2, p3, ..., p N},0<p i ≤1.0.

[0087] like Figure 3 As shown, the full-link layer network may include a full-contact layer and two full-connection (FC) layers, wherein the first FC layer may be 64-dimensional and the second FC layer may be 16-dimensional. Those skilled in the art should understand that the parameters of the above network structure are only examples and are not limited in the present disclosure.

[0088] Afterwards, a ranking result of the multiple resource modules is obtained according to the scores.

[0089] In the embodiment of the present disclosure, compared with the processing of non-long-tail words, the processing of long-tail words removes discrete features such as user-side features, resource encoding, and statistical information on the resource request side, and removes the Cartesian cross layer and DNN layer in the network structure. Through the above method, overfitting is avoided, making the resource sorting network more stable on long-tail traffic.

[0090] In some embodiments, resources may be recalled by the following method to obtain multiple resource modules.

[0091] First, the intention information of the resource request is obtained. For example, the intention analysis method in the related art can be used to perform intention analysis on the resource request to obtain the intention information of the resource request.

[0092] Next, multiple resources matching the intent information are obtained, that is, resources are recalled according to the intent information.

[0093] Since the recalled resources may be a large number of heterogeneous resources, multiple types of resources can be obtained by classifying the recalled resources. In the music search scenario, each type of resource can form a resource module.

[0094] According to the matching degree between the resources and the intent information, each type of resources is sorted to obtain a plurality of resource modules in which the internal resources are sorted.

[0095] In the disclosed embodiment, according to the matching degree between the recalled resources and the intention information of the resource request, the resource module after internal sorting of the resources can be obtained, so that the user can preferentially view the resources that meet the needs in each resource module.

[0096] The following describes a training method for a resource sorting network, which is applied to an offline system.

[0097] Figure 4 The flow chart of a method for training a resource sorting network according to an embodiment of the present disclosure is schematically shown. Figure 4 As shown, the method may include the following processing:

[0098] In step 401, the client log and server log of the online system are obtained.

[0099] Client logs and server logs can be collected in real time and cleaned and counted, or they can be collected offline. For example, they can be synchronized to a data platform and stored as offline data.

[0100] In step 402, a labeled data set and offline information are obtained based on the client log, the server log and the first resource request information stored offline, wherein the resource module in the labeled data set is labeled with a real consumption value indicating whether the resource module is consumed after exposure, and the offline features mainly include data of each resource module, resource request side data and user side data.

[0101] Among them, the first resource request information stored offline, that is, the portrait of the resource request, can be stored in the intent information library.

[0102] The client log and server log obtained from the online system, as well as the first resource request information obtained from the intent information library, are processed and summarized into annotated data sets and offline information, which can shield the differences between different resources and facilitate the unified processing of subsequent feature information extraction. In an example, the offline information may include at least one of resource module data, user data, and resource request data.

[0103] In step 403, offline feature information is acquired according to the offline information.

[0104] Feature extraction is performed on the offline information obtained in step 403 to obtain offline feature information. In one example, the obtained offline feature information includes features of each resource module, resource request features, and user features.

[0105] In step 404, a sample data set is generated according to the offline feature information and the annotated data set.

[0106] In one example, the offline information and the annotated data set may both include a resource request flag, which is generated for each resource request and is unique. Based on the offline information and the resource request flag in the annotated data set, the resource module in the annotated data set may be associated with the offline feature information to generate a sample data set.

[0107] In step 405, the resource sorting network is trained offline according to the sample data set.

[0108] In the disclosed embodiment, since the sample data in the sample data set contains the characteristics of each resource module, the problem of each resource module generating samples separately for learning and being difficult to integrate online is solved by uniformly inputting the sample data into the resource sorting network for learning.

[0109] In some implementations, the sample data in the labeled data set may include two groups of data. The first group is training target values ​​in the format of {y1, y2, y3, ..., y N},y i ∈{0, 1}, where 0 represents exposure without consumption and 1 represents consumption after exposure; the second group is the weight value of the sample {λ1, λ2, λ3, ..., λ N},0<λ i ≤1.0, where λ i Represents the learning weight of each resource module. That is, in the sample data set, each resource module is marked with a resource consumption value and a weight value. The weight value is determined according to the average exposure position of the resource module indicated in the client log. The calculation formula of the weight value is as follows:

[0110] p i =log(α+β*avgImpPos i ) (3)

[0111]

[0112] Among them, avgImpPos i is the average exposure position of the module, which is used to resist the sample bias caused by the exposure position; α and β are pre-set hyperparameters.

[0113] In some implementations, in response to the sample data in the sample data set not including long-tail words and the sample data including feature information of a user requesting resource recommendation, offline training is performed on the first resource ranking network using the sample data.

[0114] The structure of the first resource sorting network is shown in Figure 2 As shown. Sample data is input into the first resource sorting network to obtain consumption information of various resources predicted by the first resource sorting network according to the sample data; parameters of the first resource sorting network are adjusted according to the network loss function, and when the network loss is less than a set threshold or the number of iterations meets the set requirements, the trained first resource sorting network is obtained.

[0115] In one example, the network loss function is as follows:

[0116]

[0117] Among them, yi is the marked resource consumption value, p i is the predicted resource consumption value, λ i is the weight value of the i-th resource module.

[0118] In some implementations, in response to sample data in the sample data set including long-tail words, offline training is performed on the second resource ranking network using the sample data.

[0119] The structure of the second resource sorting network can be found in Figure 3 As shown. Sample data is input into the second resource sorting network to obtain the consumption information of each resource predicted by the second resource sorting network according to the sample data; the parameters of the second resource sorting network are adjusted according to the network loss function, and when the network loss is less than the set threshold or the number of iterations meets the set requirements, the trained second resource sorting network is obtained. The network loss function is, for example, as shown in formula (5).

[0120] In the disclosed embodiment, steps 101 and 102 in the resource recommendation method executed by the online system should be aligned with steps 402 and 403 in the resource ranking network training method executed by the offline system, respectively, to ensure consistency between online and offline features, so that the offline trained network can be used to score the features obtained online.

[0121] In order to implement the resource recommendation method of any embodiment of the present disclosure, the embodiment of the present disclosure also provides a resource recommendation device. Figure 5 The structure diagram of a resource recommendation device according to an embodiment of the present disclosure is schematically shown. For example, the device can be applied to a computing device. In the following description, the functions of each module of the device will be briefly described, and its detailed processing can be combined with the description of the resource recommendation method of any embodiment of the present disclosure.

[0122] like Figure 5 As shown, the device may include:

[0123] The recall unit 501 is used to obtain multiple resource modules according to the resource request, wherein the resource modules are heterogeneous and each resource module includes at least one resource;

[0124] A feature extraction unit 502 is configured to obtain online feature information, wherein the online feature information includes at least first feature information of the resource module and second feature information of the resource request;

[0125] A sorting unit 503, configured to input the online feature information into a resource sorting network to obtain sorting results of the multiple resource modules;

[0126] The recommendation unit 504 is used to recommend and display the multiple resource modules according to the ranking results of the multiple resource modules output by the resource ranking network.

[0127] In some embodiments, the device further comprises a judgment unit, which is used to judge whether the search term corresponding to the resource request is a long-tail term before inputting the online feature information into the resource ranking network to obtain the ranking results of the multiple resource modules;

[0128] The sorting unit is specifically used for:

[0129] In response to the search term not being a long-tail term, inputting the online feature information into a first resource ranking network to obtain ranking results of the plurality of resource modules;

[0130] In response to the search term being a long-tail term, the online feature information is input into a second resource ranking network to obtain ranking results of the plurality of resource modules.

[0131] In some implementations, the online feature information further includes third feature information of a user requesting resource recommendation, and the apparatus further includes a third feature extraction unit configured to obtain the third feature information according to at least one of the following:

[0132] The user's historical preference information for the resource module;

[0133] The user's historical consumption information and / or real-time consumption information for the resource module;

[0134] The user's search history information.

[0135] In some implementations, the ranking unit, in response to the search term not being a long-tail term, inputs the online feature information into the first resource ranking network to obtain ranking results of the multiple resource modules, is specifically configured to:

[0136] Converting discrete features in each type of feature information in the online feature information into fixed-length vectors, and performing a Cartesian crossover operation on each fixed-length vector to obtain a first operation result;

[0137] splicing continuous features in each type of feature information in the online feature information to obtain a splicing result, and after splicing the splicing results, performing a full connection operation to obtain a second operation result;

[0138] After splicing the first operation result and the second operation result, a full connection operation is performed to obtain the score of each resource module;

[0139] A ranking result of the multiple resource modules is obtained according to the scores.

[0140] In some embodiments, the ranking unit, in response to the search term being a long-tail term, inputs the online feature information into the second resource ranking network to obtain the ranking results of the multiple resource modules, is specifically configured to:

[0141] After concatenating the continuous features in the first feature information and the second feature information, a full connection operation is performed to obtain the score of each resource module;

[0142] A ranking result of the multiple resource modules is obtained according to the scores.

[0143] In some embodiments, the recall unit is specifically used to:

[0144] Obtaining the intent information of the resource request;

[0145] Acquire multiple resources matching the intent information;

[0146] Classifying the multiple resources to obtain multiple types of resources;

[0147] According to the matching degree between the resource and the intent information, each type of resource is sorted to obtain a plurality of resource modules.

[0148] In some embodiments, the device further includes a first feature extraction unit, configured to obtain the first feature information according to at least one of the following:

[0149] Natural language processing data of the resource module;

[0150] Historical consumption information and / or real-time consumption information of the resource module;

[0151] Text matching information between the resource request and the resource module;

[0152] A profile of the resource module and a profile of each resource in the resource module.

[0153] In some embodiments, the device further includes a second feature extraction unit, configured to obtain the second feature information according to at least one of the following:

[0154] The intent information of the resource request;

[0155] natural language processing data of the resource request;

[0156] The historical consumption information and / or real-time consumption information of each resource module under the resource request.

[0157] Figure 6The structure diagram of a training device for a resource sorting network according to an embodiment of the present disclosure is schematically shown. For example, the device can be applied to a computing device. In the following description, the functions of each module of the device will be briefly described, and its detailed processing can be combined with the description of the resource recommendation method of any embodiment of the present disclosure.

[0158] like Figure 6 As shown, the device comprises:

[0159] The log acquisition unit 601 is used to acquire the client log and server log of the online system;

[0160] The data acquisition unit 602 is used to obtain a labeled data set and offline information according to the client log, the server log and the first resource request information stored offline, wherein the offline information at least includes resource module information and second resource request information, and the resource module in the labeled data set is labeled with a consumption real value indicating whether the resource module is consumed after exposure;

[0161] A feature acquisition unit 603, configured to acquire offline feature information according to the offline information, wherein the offline feature information at least includes the fourth feature information of the resource module and the fifth feature information of the resource request;

[0162] A sample generating unit 604, configured to generate a sample data set according to the offline feature information and the annotated data set;

[0163] The training unit 605 is used to perform offline training on the resource sorting network according to the sample data set.

[0164] In some embodiments, the training unit is specifically used for:

[0165] In response to the sample data in the sample data set not including long-tail words, and the sample data including feature information of a user requesting resource recommendation, offline training a first resource ranking network using the sample data;

[0166] In response to the sample data in the sample data set including long-tail words, offline training is performed on a second resource ranking network using the sample data.

[0167] In some implementations, the resource modules in the annotated data set are also annotated with weight values ​​of the resource modules, wherein the weight values ​​are determined according to an average exposure position of the resource modules indicated in the client log.

[0168] In some implementations, the offline information and the annotated data set both include a resource request flag, the resource request flag is generated for each resource request and is unique, and the sample generation unit is specifically used to:

[0169] According to the offline information and the resource request flag in the annotated data set, the resource module in the annotated data set is associated with the offline feature information to generate sample data.

[0170] It should be noted that, although several units / modules or sub-units / modules of the apparatus for generating summaries for audio files are mentioned in the above detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above may be embodied in one unit / module. Conversely, the features and functions of one unit / module described above may be further divided to be embodied by multiple units / modules.

[0171] The present disclosure also provides a computer-readable storage medium. Figure 7 As shown, the storage medium stores a computer program 701, and when the computer program 701 is executed by a processor, the resource recommendation method of any embodiment of the present disclosure can be executed.

[0172] An embodiment of the present disclosure further provides a computing device, which may include a memory and a processor, wherein the memory is used to store computer instructions executable on the processor, and the processor is used to implement the resource recommendation method of any embodiment of the present disclosure when executing the computer instructions.

[0173] Figure 8 An example of a structure of the computing device is shown in FIG. Figure 8 As shown, the computing device 80 may include, but is not limited to: a processor 81 , a memory 82 , and a bus 83 connecting different system components (including the memory 82 and the processor 81 ).

[0174] The memory 82 stores computer instructions, which can be executed by the processor 81, so that the processor 81 can execute the resource recommendation method of any embodiment of the present disclosure. The memory 82 may include a random access memory unit RAM 821, a cache memory unit 822 and / or a read-only memory unit ROM 823. The memory 82 may also include: a program tool 825 having a set of program modules 824, the program modules 824 include but are not limited to: an operating system, one or more application programs, other program modules and program data, and one or more combinations of these program modules may include the implementation of a network environment.

[0175] The bus 83 may include, for example, a data bus, an address bus, and a control bus. The computing device 80 may also communicate with an external device 85 through an I / O interface 84. The external device 85 may be, for example, a keyboard, a Bluetooth device, etc. The computing device 80 may also communicate with one or more networks through a network adapter 88. For example, the network may be a local area network, a wide area network, a public network, etc. Figure 8 As shown, the network adapter 88 can also communicate with other modules of the computing device 80 via the bus 83 .

[0176] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0177] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A resource recommendation method, characterized in that: Applied to online systems, including: Acquire multiple resource modules according to the resource request, wherein the resource modules are heterogeneous and each resource module includes at least one resource; Acquire online characteristic information, where the online characteristic information at least includes first characteristic information of the resource module and second characteristic information of the resource request; Inputting the online feature information into a resource sorting network to obtain sorting results of the plurality of resource modules, including: organizing the first feature information of all resource modules according to an input format of the resource sorting network, splicing them according to a fixed order of resource modules, and then inputting them into the resource sorting network; According to the sorting results of the multiple resource modules output by the resource sorting network, recommend and display the multiple resource modules; Wherein, the obtaining of multiple resource modules according to the resource request includes: Obtaining the intent information of the resource request; Acquire multiple resources matching the intent information; Classifying the multiple resources to obtain multiple types of resources; According to the matching degree between the resource and the intent information, each type of resource is sorted to obtain a plurality of resource modules.

2. The method according to claim 1, characterized in that Before inputting the online feature information into the resource ranking network to obtain the ranking results of the multiple resource modules, the method further includes determining whether the search term corresponding to the resource request is a long-tail term; The step of inputting the online feature information into a resource sorting network to obtain sorting results of the plurality of resource modules includes: In response to the search term not being a long-tail term, inputting the online feature information into a first resource ranking network to obtain ranking results of the plurality of resource modules; In response to the search term being a long-tail term, the online feature information is input into a second resource ranking network to obtain ranking results of the plurality of resource modules.

3. The method according to claim 2, characterized in that In response to the search term not being a long-tail term, the online feature information further includes third feature information of the user requesting resource recommendation, and the method further includes obtaining the third feature information according to at least one of the following: The user's historical preference information for the resource module; The user's historical consumption information and / or real-time consumption information for the resource module; The user's search history information.

4. The method according to claim 2, characterized in that: In response to the search term not being a long-tail term, inputting the online feature information into a first resource ranking network to obtain ranking results of the multiple resource modules includes: Converting discrete features in each type of feature information in the online feature information into fixed-length vectors, and performing a Cartesian crossover operation on each fixed-length vector to obtain a first operation result; splicing continuous features in each type of feature information in the online feature information to obtain a splicing result, and after splicing the splicing results, performing a full connection operation to obtain a second operation result; After concatenating the first operation result and the second operation result, a full connection operation is performed to obtain the score of each resource module; A ranking result of the multiple resource modules is obtained according to the scores.

5. The method according to claim 2, characterized in that: In response to the search term being a long-tail term, inputting the online feature information into a second resource sorting network to obtain sorting results of the multiple resource modules includes: After concatenating the continuous features in the first feature information and the second feature information, a full connection operation is performed to obtain the score of each resource module; A ranking result of the multiple resource modules is obtained according to the scores.

6. The method according to any one of claims 1 to 5, characterized in that: The method further includes obtaining the first feature information according to at least one of the following: Natural language processing data of the resource module; Historical consumption information and / or real-time consumption information of the resource module; Text matching information between the resource request and the resource module; A profile of the resource module and a profile of each resource in the resource module.

7. The method according to any one of claims 1 to 5, characterized in that: The method further includes obtaining the second feature information according to at least one of the following: The intent information of the resource request; natural language processing data of the resource request; The historical consumption information and / or real-time consumption information of each resource module under the resource request.

8. A method for training a resource sorting network, characterized in that: Used to train a resource sorting network that sorts using the method described in any one of claims 1 to 7, wherein the training method is applied to an offline system and comprises: Obtain client logs and server logs of online systems; Obtaining a labeled data set and offline information according to the client log, the server log, and the first resource request information stored offline, wherein the resource module in the labeled data set is labeled with a consumption real value indicating whether the resource module is consumed after being exposed; Acquire offline feature information according to the offline information; Generate a sample data set according to the offline feature information and the annotated data set; Offline training of the resource sorting network is performed according to the sample data set, including: The sample data is spliced ​​in a fixed order of resource modules. The resource sorting network is used to sort multiple resource modules obtained by the online system according to resource requests. The first feature information of all resource modules is organized according to the input format of the resource sorting network, and is spliced ​​in the fixed order of the resource modules and then input into the resource sorting network.

9. The method according to claim 8, characterized in that The offline training of the resource sorting network according to the sample data set includes: In response to the sample data in the sample data set not including long-tail words, and the sample data including feature information of a user requesting resource recommendation, offline training a first resource ranking network using the sample data; In response to the sample data in the sample data set including long-tail words, offline training is performed on a second resource ranking network using the sample data.

10. The method according to claim 8 or 9, characterized in that: The resource modules in the annotated data set are also annotated with weight values ​​of the resource modules, wherein the weight values ​​are determined according to an average exposure position of the resource modules indicated in the client log.

11. The method according to claim 8 or 9, characterized in that: The offline information and the annotated data set both include a resource request flag, and the resource request flag is generated for each resource request and is unique. The generating of the sample data set according to the offline feature information and the annotated data set includes: According to the offline information and the resource request flag in the annotated data set, the resource modules in the annotated data set are associated with the offline feature information to generate a sample data set.

12. A resource recommendation device, characterized in that: Applied to online systems, including: A recall unit, configured to obtain a plurality of resource modules according to a resource request, wherein the resource modules are heterogeneous and each of the resource modules includes at least one resource; A feature extraction unit, which obtains online feature information, wherein the online feature information at least includes first feature information of the resource module and second feature information of the resource request; A sorting unit, used for inputting the online feature information into a resource sorting network to obtain a sorting result of the plurality of resource modules, comprising: organizing the first feature information of all resource modules according to an input format of the resource sorting network, splicing them according to a fixed order of resource modules and then inputting them into the resource sorting network; A recommendation unit, configured to recommend and display the multiple resource modules according to the ranking results of the multiple resource modules output by the resource ranking network; The recall unit is specifically used for: Obtaining the intent information of the resource request; Acquire multiple resources matching the intent information; Classifying the multiple resources to obtain multiple types of resources; According to the matching degree between the resource and the intent information, each type of resource is sorted to obtain a plurality of resource modules.

13. The device according to claim 12, characterized in that The device further comprises a judgment unit, which is used to judge whether the search term corresponding to the resource request is a long-tail term before inputting the online feature information into the resource ranking network to obtain the ranking results of the plurality of resource modules; The sorting unit is specifically used for: In response to the search term not being a long-tail term, inputting the online feature information into a first resource ranking network to obtain ranking results of the plurality of resource modules; In response to the search term being a long-tail term, the online feature information is input into a second resource ranking network to obtain ranking results of the plurality of resource modules.

14. The device according to claim 13, characterized in that The online feature information further includes third feature information of the user requesting resource recommendation, and the apparatus further includes a third feature extraction unit, configured to obtain the third feature information according to at least one of the following: The user's historical preference information for the resource module; The user's historical consumption information and / or real-time consumption information for the resource module; The user's search history information.

15. The device according to claim 14, characterized in that When the ranking unit is used to input the online feature information into the first resource ranking network in response to the search word not being a long-tail word to obtain the ranking results of the multiple resource modules, it is specifically used to: Converting discrete features in each type of feature information in the online feature information into fixed-length vectors, and performing a Cartesian crossover operation on each fixed-length vector to obtain a first operation result; splicing continuous features in each type of feature information in the online feature information to obtain a splicing result, and after splicing the splicing results, performing a full connection operation to obtain a second operation result; After concatenating the first operation result and the second operation result, a full connection operation is performed to obtain the score of each resource module; A ranking result of the multiple resource modules is obtained according to the scores.

16. The device according to claim 13, characterized in that The sorting unit, when used to input the online feature information into the second resource sorting network in response to the search word being a long-tail word to obtain the sorting results of the multiple resource modules, is specifically used to: After concatenating the continuous features in the first feature information and the second feature information, a full connection operation is performed to obtain the score of each resource module; A ranking result of the multiple resource modules is obtained according to the scores.

17. The device according to any one of claims 12 to 16, characterized in that The device further includes a first feature extraction unit, configured to obtain the first feature information according to at least one of the following: Natural language processing data of the resource module; Historical consumption information and / or real-time consumption information of the resource module; Text matching information between the resource request and the resource module; A profile of the resource module and a profile of each resource in the resource module.

18. The device according to any one of claims 12 to 16, characterized in that The device further includes a second feature extraction unit, configured to obtain the second feature information according to at least one of the following: The intent information of the resource request; natural language processing data of the resource request; The historical consumption information and / or real-time consumption information of each resource module under the resource request.

19. A training device for a resource sorting network, characterized in that: Used to train a resource sorting network that sorts using the method described in any one of claims 1 to 7, wherein the training device is applied to an offline system, comprising: Log acquisition unit, used to obtain client logs and server logs of online systems; A data acquisition unit, configured to obtain a labeled data set and offline information according to the client log, the server log, and the first resource request information stored offline, wherein the offline information at least includes resource module information and second resource request information, and the resource module in the labeled data set is labeled with a consumption real value indicating whether the resource module is consumed after being exposed; A feature acquisition unit, configured to acquire offline feature information according to the offline information, wherein the offline feature information at least includes fourth feature information of the resource module and fifth feature information of the resource request; A sample generating unit, configured to generate a sample data set according to the offline feature information and the annotated data set; A training unit, used for performing offline training on the resource sorting network according to the sample data set, comprising: The sample data is spliced ​​in a fixed order of resource modules. The resource sorting network is used to sort multiple resource modules obtained by the online system according to resource requests. The first feature information of all resource modules is organized according to the input format of the resource sorting network, and is spliced ​​in the fixed order of the resource modules and then input into the resource sorting network.

20. The device according to claim 19, characterized in that The training unit is specifically used for: In response to the sample data in the sample data set not including long-tail words, and the sample data including feature information of a user requesting resource recommendation, offline training a first resource ranking network using the sample data; In response to the sample data in the sample data set including long-tail words, offline training is performed on a second resource ranking network using the sample data.

21. The device according to claim 19 or 20, characterized in that The resource modules in the annotated data set are also annotated with weight values ​​of the resource modules, wherein the weight values ​​are determined according to an average exposure position of the resource modules indicated in the client log.

22. The device according to claim 19 or 20, characterized in that The offline information and the annotated data set both include a resource request flag, which is generated for each resource request and is unique. The sample generation unit is specifically used to: According to the offline information and the resource request flag in the annotated data set, the resource module in the annotated data set is associated with the offline feature information to generate sample data.

23. A computing device, characterized in that include: processor; as well as A memory having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by the processor, implement the method according to any one of claims 1 to 7, or the method according to any one of claims 8 to 11.

24. A computer-readable storage medium, characterized in that: The method comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 7, or any one of claims 8 to 11.

Images