WO2021164643A1 - 作品推荐方法及服务器 - Google Patents

作品推荐方法及服务器 Download PDF

Info

Publication number
WO2021164643A1
WO2021164643A1 PCT/CN2021/076195 CN2021076195W WO2021164643A1 WO 2021164643 A1 WO2021164643 A1 WO 2021164643A1 CN 2021076195 W CN2021076195 W CN 2021076195W WO 2021164643 A1 WO2021164643 A1 WO 2021164643A1
Authority
WO
WIPO (PCT)
Prior art keywords
works
multimedia
work
account
recommendation
Prior art date
Application number
PCT/CN2021/076195
Other languages
English (en)
French (fr)
Inventor
文浩丞
曾钢
赵彦宾
Original Assignee
北京达佳互联信息技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京达佳互联信息技术有限公司 filed Critical 北京达佳互联信息技术有限公司
Priority to JP2022547740A priority Critical patent/JP2023512692A/ja
Publication of WO2021164643A1 publication Critical patent/WO2021164643A1/zh
Priority to US17/891,465 priority patent/US20220398277A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/43—Querying
    • G06F16/435—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/43—Querying
    • G06F16/435—Filtering based on additional data, e.g. user or group profiles
    • G06F16/437—Administration of user profiles, e.g. generation, initialisation, adaptation, distribution
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/43—Querying
    • G06F16/438—Presentation of query results
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation

Definitions

  • the present disclosure relates to the field of information technology, and in particular to a method and server for recommending works.
  • feeds combines several information sources that users actively subscribe to form a content aggregator to help users continuously obtain the latest feed content.
  • the present disclosure provides a method and server for recommending works.
  • the technical solutions of the present disclosure are as follows:
  • a method for recommending a work comprising: receiving a recommendation request sent by a login account of an application, wherein the recommendation request is used to request to display a multimedia work, and the multimedia work is The login account is based on the multimedia works published by the associated account of the application; in response to the recommendation request, a first candidate set of works of each of a plurality of types is obtained, and the first candidate set of works includes the Multimedia works published by an associated account; screening each of the first candidate sets of works, and summarizing the screening results into a second candidate set of works, the second candidate set of works including the multiple types of multimedia works; The multiple types of multimedia works in the second work candidate set are sorted, and the multimedia works are recommended to the login account according to the sorting result.
  • the sorting of the multiple types of multimedia works in the second work candidate set includes: according to participation and recommended guidance information set by an application platform, comparing the second work candidate set The multimedia works are sorted, the participation degree is used to indicate the positive feedback operation or the negative feedback operation performed by the account on the historical multimedia works, and the recommendation guidance information includes information used to indicate the degree of recommendation of the historical multimedia works by the application platform. At least one item of recommendation information or guidance information used to prompt the account to perform a positive feedback operation on historical multimedia works.
  • the sorting of the multimedia works in the second work candidate set according to the participation degree and the recommended guidance information set by the application platform includes: collecting the second work candidate sets of the multiple types of works The multimedia works are input into a mixed ranking model to obtain the sorting sequences of the multiple types of multimedia works, and the mixed ranking model is obtained by training according to the participation degree and the recommended guidance information set by the application platform.
  • a method for training a work recommendation hybrid ranking model includes: acquiring a plurality of types of sample sets, the sample sets including positive samples and negative samples, and the positive samples are Refers to historical multimedia works that were clicked by the account after being displayed, and the negative sample refers to historical multimedia works that were not clicked on by the account after being displayed; determined according to participation and the recommended guidance information set by the application platform
  • the ranking score of each of the positive samples in the sample set, the participation degree is used to indicate the positive feedback operation or the negative feedback operation performed by the account on the historical multimedia works
  • the recommended guidance information includes information used to indicate the application At least one of the recommendation information of the platform's recommendation for historical multimedia works or the guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia works; according to the ranking score and the ranking score of each positive sample in the sample set In the sample set, a hybrid ranking model is trained.
  • the training a hybrid ranking model based on the ranking score of each positive sample in the sample set and the sample set includes: generating and ranking score for each positive sample, etc.
  • a number of target positive samples based on the sample set and target positive samples, a positive sample probability determination model is trained, and the positive sample probability determination model is used to determine the probability of the target positive sample; based on the probability of the target positive sample And the probability that each sample in the sample set is the positive sample, and train the hybrid ranking model.
  • the determining the ranking score of each of the positive samples in the sample set according to the participation degree and the recommendation guidance information set by the application platform includes: obtaining each account for each of the The positive feedback operation performed by the positive sample and its weight; the negative feedback operation performed by each account on each positive sample and its weight; based on the obtained positive feedback operation and its weight and negative feedback Operation and its weight to determine the degree of participation of each account for each positive sample; based on the recommended guidance information set by the application platform for each positive sample, determine each account for each positive sample The weights of participation of the positive samples; and based on the participation of each of the accounts for each of the positive samples and the weights thereof, the ranking score of each of the positive samples in the sample set is determined.
  • the determining the participation degree of each of the accounts for each of the positive samples based on the acquired positive feedback operations and their weights and the negative feedback operations and their weights includes: The feedback operation, in response to determining that the current feedback operation belongs to the low-frequency feedback operation, adjust the weight of the current feedback operation to the ratio of the occurrence frequency of the target high-frequency feedback operation to the occurrence frequency of the low-frequency feedback operation, and the target
  • the frequency of occurrence of high-frequency feedback operations refers to the average frequency of occurrence of all high-frequency feedback operations in all feedback operations currently acquired; each feedback operation and its adjusted weight are used to determine each account pair for each positive sample Participation.
  • a work recommendation device comprising: a receiving module configured to receive a recommendation request sent by a login account of an application, wherein the recommendation request is used to request display of multimedia A work, the multimedia work is a multimedia work published by the login account based on the associated account of the application; the obtaining module is configured to obtain the first work of each of the multiple types in response to the recommendation request A candidate set, the first candidate set of works includes multimedia works published by the associated account; a screening summary module configured to screen each of the first candidate sets of works, and summarize the screening results into a second candidate set of works, The second candidate work set includes the multiple types of multimedia works; the sorting module is configured to rank the multiple types of multimedia works in the second work candidate set, and according to the sorting result, to the Log in to your account to recommend multimedia works.
  • the sorting module is configured to sort the multimedia works in the second candidate set of works according to the participation degree and the recommendation guidance information set by the application platform, and the participation degree is used to indicate account pairing.
  • the recommendation guide information includes recommendation information used to indicate the degree of recommendation of the application platform to the historical multimedia work or used to prompt the account to perform a positive feedback on the historical multimedia work At least one item in the guidance information of the feedback operation.
  • the sorting module includes: inputting the multiple types of multimedia works in the second work candidate set into a mixed sorting model to obtain a sorting sequence of the multiple types of multimedia works, and the mixed sorting The model is obtained by training according to the participation degree and the recommended guidance information set by the application platform.
  • the ranking module includes: a training sub-module configured to train a hybrid ranking model according to the participation degree and the recommendation guidance information set by the application platform, and the hybrid ranking model is used for Determine the sorting sequence of the multimedia works according to the degree of participation and the recommended guidance information; the sorting sub-module is configured to input the multiple types of multimedia works in the second work candidate set into the training sub-module for training The hybrid ranking model is used to obtain the ranking sequence of the multiple types of multimedia works in the second work candidate set.
  • the training sub-module includes: an acquisition unit configured to acquire a plurality of types of sample sets, the sample sets include positive samples and negative samples, and the positive samples refer to those that are displayed by the account after being displayed.
  • Historical multimedia works that perform a click operation the negative sample refers to historical multimedia works that the account did not perform a click operation after being displayed;
  • the ranking score of each of the positive samples in the sample set, the participation degree is used to indicate the positive feedback operation or the negative feedback operation performed by the account on the historical multimedia works, and the recommended guidance information includes information used to indicate the application platform At least one of the recommendation information for the recommendation degree of historical multimedia works or the guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia works;
  • the ranking score and the sample set are used to train a hybrid ranking model.
  • the training unit is configured to: generate a target positive sample of the same number as its ranking score for each positive sample; and train a positive sample probability determination model based on the sample set and the target positive sample
  • the positive sample probability determination model is used to determine the probability of the target positive sample; based on the probability of the target positive sample and the probability that each sample in the sample set is the positive sample, the hybrid ranking model is trained .
  • the determining unit is configured to: obtain a positive feedback operation performed by each account for each positive sample and its weight; obtain each account for each positive sample.
  • the negative feedback operation performed by the sample and its weight based on the acquired positive feedback operation and its weight, and the negative feedback operation and its weight, determine the participation of each account for each positive sample; based on all
  • the application platform sets the recommended guidance information for each of the positive samples to determine the weight of each account’s participation in each of the positive samples; The degree of participation and its weight determine the ranking score of each positive sample in the sample set.
  • the determining unit is configured to: for each feedback operation, in response to determining that the current feedback operation belongs to a low-frequency feedback operation, adjust the weight of the current feedback operation to a target high-frequency feedback operation The ratio of the occurrence frequency of the low-frequency feedback operation to the occurrence frequency of the low-frequency feedback operation.
  • the occurrence frequency of the target high-frequency feedback operation refers to the average occurrence frequency of all high-frequency feedback operations in all feedback operations currently acquired; according to each feedback operation and The adjusted weight determines the participation degree of each account for each positive sample.
  • an apparatus for training a work recommendation hybrid ranking model including: an obtaining unit configured to obtain a plurality of types of sample sets, the sample sets including positive samples and negative samples
  • the positive sample refers to a historical multimedia work that was clicked by the account after being displayed, and the negative sample refers to a historical multimedia work that was not clicked on by the account after being displayed;
  • the determining unit is configured to be based on the degree of participation Determining the ranking score of each positive sample in the sample set with the recommended guidance information set by the application platform, and the degree of participation is used to indicate the positive feedback operation or the negative feedback operation performed by the account on the historical multimedia works
  • the recommended guidance information includes at least one of recommendation information used to indicate the degree of recommendation of the historical multimedia work by the application platform or guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia work
  • the training unit is configured To train a hybrid ranking model according to the ranking score of each positive sample in the sample set and the sample set.
  • the training unit is configured to: generate a target positive sample of the same number as its ranking score for each positive sample; and train a positive sample probability determination model based on the sample set and the target positive sample
  • the positive sample probability determination model is used to determine the probability of the target positive sample; based on the probability of the target positive sample and the probability that each sample in the sample set is the positive sample, the hybrid ranking model is trained .
  • the determining unit is configured to: obtain a positive feedback operation performed by each account for each positive sample and its weight; obtain each account for each positive sample.
  • the negative feedback operation performed by the sample and its weight based on the acquired positive feedback operation and its weight, and the negative feedback operation and its weight, determine the participation of each account for each positive sample; based on all
  • the application platform sets the recommended guidance information for each of the positive samples to determine the weight of each account’s participation in each of the positive samples; The degree of participation and its weight determine the ranking score of each positive sample in the sample set.
  • the determining unit is configured to: for each feedback operation, in response to determining that the current feedback operation belongs to a low-frequency feedback operation, adjust the weight of the current feedback operation to a target high-frequency feedback operation The ratio of the occurrence frequency of the low-frequency feedback operation to the occurrence frequency of the low-frequency feedback operation.
  • the occurrence frequency of the target high-frequency feedback operation refers to the average occurrence frequency of all high-frequency feedback operations in all feedback operations currently acquired; according to each feedback operation and The adjusted weight determines the participation degree of each account for each positive sample.
  • a server including:
  • a memory for storing executable instructions of the processor
  • the processor is configured to execute the instruction to implement the following operations: receiving a recommendation request sent by the login account of the application program, wherein the recommendation request is used to request to display a multimedia work, and the multimedia work is the The login account is based on the multimedia works published by the associated account of the application; in response to the recommendation request, a first candidate set of works of each of a plurality of types is obtained, and the first candidate set of works includes the associated account Multimedia works released; screening each of the first candidate sets of works, and summarizing the screening results into a second candidate set of works, the second candidate set of works including the multiple types of multimedia works; The multiple types of multimedia works in the work candidate set are sorted, and the multimedia works are recommended to the login account according to the sorting result.
  • a storage medium which is executed by a processor of a server in response to instructions in the storage medium, so that the server can perform the following operations: receiving a recommendation request sent by a login account of an application program, Wherein, the recommendation request is used to request to display a multimedia work, and the multimedia work is a multimedia work published by the login account based on the associated account of the application; in response to the recommendation request, each of the multiple types is obtained
  • the first work candidate set of the type, the first work candidate set includes the multimedia works published by the associated account; each of the first work candidate sets is screened, and the screening results are aggregated into the second work candidate set, the
  • the second candidate set of works includes the multiple types of multimedia works; the multiple types of multimedia works in the second candidate set are sorted, and the multimedia works are recommended to the login account according to the ranking result.
  • Fig. 1 is a flowchart of a method for recommending works according to an embodiment of the present disclosure.
  • Fig. 2 is a flowchart of training a hybrid ranking model according to an embodiment of the present disclosure.
  • Fig. 3 is a flowchart of determining the ranking score of each positive sample in a sample set according to an embodiment of the present disclosure.
  • Fig. 4 is a block diagram of an apparatus for recommending works according to an embodiment of the present disclosure.
  • Fig. 5 is a block diagram of another device for recommending works according to an embodiment of the present disclosure.
  • Fig. 6 is a block diagram of another device for recommending works according to an embodiment of the present disclosure.
  • Fig. 7 is a block diagram of a server according to an embodiment of the present disclosure.
  • Fig. 8 is a block diagram of a device suitable for a method for recommending works according to an embodiment of the present disclosure.
  • FIG. 1 is a flowchart of a method for recommending works according to an embodiment of the present disclosure. As shown in FIG. 1, the method for recommending works includes the following contents:
  • a recommendation request sent by the login account of the application is received, where the recommendation request is used to request the display of multimedia works on the target page, and the target page is used to display the association account that has established a social relationship with the login account through the application. Published multimedia works.
  • the multimedia work is a multimedia work published by the login account based on the associated account of the application program.
  • the login account and the associated account establish a social relationship based on the application program.
  • the target page may include but is not limited to the follow page, and may also include the same city page.
  • Multimedia works can include, but are not limited to, live broadcasts, short videos, and other works.
  • the application may send a recommendation request for the login account to the server.
  • the application may include, but is not limited to, an APP that publishes multimedia works.
  • the associated account that has a social relationship with the login account may include the follow account of the login account.
  • the associated account that has established a social relationship with the login account may include an account located in the same city as the login account.
  • a first candidate set of works belonging to various types of multimedia works published by the associated account is obtained from the works library.
  • This S102 is a possible implementation manner of obtaining a first candidate set of works of each of a plurality of types in response to the recommendation request, where the first candidate set of works includes multimedia works published by the associated account.
  • the respective types refer to each of the multiple types.
  • the server may, in response to the recommendation request, obtain the first candidate set of works from the work library, where each type of candidate set of first works may include, but is not limited to, the first candidate set of live broadcast, Candidate set of the first works in the short video category.
  • the first candidate set of works may also be located at other addresses.
  • the first candidate set of works is located in the local storage of the server, which is not specifically limited in the embodiment of the present disclosure.
  • This S103 is a possible implementation of screening each of the first candidate sets of works, and summarizing the screening results into a second candidate set of works, the second candidate set of works including the multiple types of multimedia works.
  • the screening can be based on server processing parameters or other screening rules, which is not specifically limited in the embodiment of the present disclosure.
  • the candidate sets of the first works may be screened separately based on the performance of the server, such as server processing parameters, and the screening results are summarized into the candidate set of the second works.
  • the multimedia works in the second work candidate set are sorted, and the multimedia works are recommended to the client according to the sorting result.
  • the S104 is a process of sorting the multiple types of multimedia works in the second work candidate set, and recommending the multimedia works to the login account according to the sorting result.
  • the multimedia works can be sent to the client with the login account.
  • the recommendation process may be to display multimedia works on the client according to the sorted results. For example, multimedia works with the highest ranking results are displayed first.
  • the multimedia works in the second candidate set can be sorted according to the degree of participation and the recommendation guidance information set by the application platform, that is, the multimedia works in the second candidate set are sorted using a unified standard. It is helpful to improve the accuracy of the sorting results.
  • Positive feedback operations may include, but are not limited to, viewing operations, like operations, following operations, comment operations, etc.
  • negative feedback operations may include, but are not limited to, including reporting operations.
  • the recommended guidance information may include at least one of recommendation information used to indicate the degree of recommendation of the historical multimedia work by the application platform and guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia work. That is to say, there are three situations: the recommendation guide information may include recommendation information used to indicate the degree of recommendation of the historical multimedia works by the application platform. The recommended guidance information may include guidance information for prompting the account to perform a positive feedback operation on the historical multimedia works. The recommended guidance information may include recommendation information used to indicate the degree of recommendation of the historical multimedia work by the application platform and guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia work.
  • sorting the multimedia works in the second work candidate set according to the participation degree and the recommendation guide information set by the application platform may include: collecting the second work candidate sets of the multiple types of multimedia The work is input into a hybrid ranking model to obtain the ranking sequence of the multiple types of multimedia works, and the hybrid ranking model is trained according to the participation degree and the recommended guidance information set by the application platform.
  • the training process may be: training a hybrid ranking model according to the participation degree and recommendation guidance information set by the application platform, and the hybrid ranking model is used to determine the ranking sequence of the multimedia works according to the participation degree and the recommendation guidance information.
  • the training process can be completed in advance, and the trained hybrid ranking model is directly used when ranking the second work candidate set.
  • the training process can also be executed when the second work candidate set needs to be sorted.
  • the embodiment of the present disclosure by receiving a recommendation request sent by the login account of the application, and in response to the recommendation request, obtaining the first candidate set of works belonging to various types of multimedia works published by the associated account from the works library, and then at least based on After the server processing parameters screen each first work candidate set, the screening results are summarized into the second work candidate set, and finally the multimedia works in the second work candidate set are sorted, and the multimedia works are recommended to the client according to the sorting results , Because the embodiment of the present disclosure is to sort the works in the second candidate set of works after screening the first candidate set containing multiple types, that is, the embodiment of the present disclosure is to uniformly sort the works of each type , Which helps to improve the accuracy of the sorting results, thereby improving the accuracy of recommending multimedia works based on the sorting results.
  • FIG. 2 is a flowchart of training a hybrid ranking model according to an embodiment of the present disclosure. As shown in Figure 2, the process of training the hybrid ranking model may include the following:
  • a sample set of multiple types of multimedia works is obtained.
  • the sample set includes a positive sample and a negative sample.
  • This S201 is a process of obtaining multiple types of sample sets.
  • the sample sets include samples, and each sample can be any type of multimedia work of multiple types.
  • the positive sample refers to a historical multimedia work whose click operation is performed by the account after being displayed
  • the negative sample refers to a historical multimedia work whose click operation is not performed by the account after being displayed.
  • the presentation process is displayed on the target page, or may be displayed on other pages, which is not limited in the embodiment of the present disclosure.
  • multiple types of multimedia works may include, but are not limited to, live broadcasts, short videos, and other works.
  • the above-mentioned presentation log includes a user identification (userId) and a work (Item).
  • userId user identification
  • Item work
  • the item of the account is a sample.
  • the item being clicked by the account the item is a positive sample.
  • the item being not clicked by the account the item is a negative sample.
  • display logs of multiple types of multimedia works can be obtained, and in response to a click operation performed by the account after the corresponding multimedia works are determined to be displayed according to the display logs, a positive sample is generated, and the label is recorded as 1. In response to determining according to the presentation log that the account does not perform a click operation after the corresponding multimedia work is displayed, a negative sample is generated, and the label is recorded as 0.
  • the ranking score of each positive sample in the sample set is determined according to the degree of participation and the recommendation guidance information set by the application platform.
  • the participation degree of the account can be determined based on the positive feedback operation and the negative feedback operation and its weight performed by the account on the historical multimedia works.
  • the recommended guidance information set by the application platform can be set by the application platform based on ecological factors, for example Ecological factors can include but are not limited to the degree of traffic inclusiveness, etc., and can also be set based on other factors.
  • determining the ranking score of each positive sample in the sample set can include:
  • S2021 obtain the positive feedback operation performed by each account on each positive sample and its weight, and the negative feedback operation performed by each account on each positive sample and its weight, and perform each account on each positive sample.
  • the positive feedback operations performed by the positive samples and their weights and the negative feedback operations performed by each account on each positive sample and their weights determine the participation of each account on each positive sample.
  • the process of determining the degree of participation in S2021 is a process of determining the degree of participation of each account for each positive sample based on the acquired positive feedback operations and their weights and negative feedback operations and their weights.
  • positive feedback operations may include but are not limited to viewing operations, like operations, following operations, comment operations, etc.
  • negative feedback operations may include, but are not limited to, reporting operations.
  • the weight of the negative feedback operation or the weight of the negative feedback operation performed by the account on each positive sample can be determined by the retention attribution algorithm, and then the weighting operation is performed according to each feedback operation and its weight , Get the participation degree of each account for each positive sample.
  • the behavior is a feedback operation, which may be a positive feedback operation or a negative feedback operation.
  • the weight of the current feedback operation is adjusted to the ratio of the occurrence frequency of the target high-frequency feedback operation to the occurrence frequency of the low-frequency feedback operation.
  • the frequency of occurrence of the target high-frequency feedback operation refers to the average frequency of occurrence of all high-frequency feedback operations in all feedback operations currently acquired; each feedback operation and its adjusted weight are used to determine each account for each account. State the participation of the positive sample.
  • the weight of the current behavior is adjusted to the ratio of the occurrence frequency of the target high-frequency behavior to the occurrence frequency of the low-frequency behavior, and according to each behavior and its adjustment
  • the latter weight determines the participation degree of each account for each positive sample.
  • the occurrence frequency of the target high-frequency behavior refers to the average occurrence frequency of all high-frequency behaviors among all the behaviors currently acquired.
  • all the behaviors currently acquired are likes and comments, where likes are high-frequency behaviors, comments are low-frequency behaviors, the pre-statistical frequency of likes is 0.1, and the pre-statistical frequency of comments is 0.001. Adjust the weight of the comment to 100. It should be noted that the behavior and values involved in this example are only an example, and can be adjusted as needed in actual applications.
  • the recommended guidance information set by the application platform for each positive sample is used to determine the weight of each account's participation in each positive sample.
  • This S2022 is a process of determining the weight of participation of each of the accounts for each of the positive samples based on the recommended guidance information set by the application platform for each of the positive samples.
  • the account accepts different types of multimedia works differently. For example, the account accepts short videos higher than the live broadcast. At this time, the weight of the account's participation in short videos can be increased.
  • the S2023 is a process of determining the ranking score of each positive sample in the sample set based on the participation degree and weight of each positive sample by each account.
  • the ranking score of each positive sample in the sample set can be obtained according to the participation of each account in each positive sample and its weight.
  • the ranking score of each positive sample in the sample set is determined according to the degree of participation and the recommendation guidance information set by the application platform, so that different types of multimedia works can measure the ranking information based on a unified standard, which is beneficial to improve The accuracy of the trained hybrid ranking model.
  • the degree of participation and the degree of recommendation set by the platform can include multiple dimensions of information, this information can well describe the characteristics of the application scenario, which can further improve the accuracy of the trained hybrid ranking model, that is, further improve the The hybrid ranking model determines the accuracy of the ranking sequence.
  • a new sample set is generated according to the ranking score and the sample set of each positive sample in the sample set, and a hybrid ranking model is trained based on the new sample set.
  • This S203 is a process of training the hybrid ranking model according to the ranking score of each positive sample in the sample set and the sample set.
  • This process may not generate a new sample set, but directly use the sample set and the generated positive sample, and the generated positive sample may be referred to as the target positive sample.
  • the training process of the hybrid ranking model may include: generating a target positive sample equal to its ranking score for each positive sample; training a positive sample probability determination model based on the sample set and the target positive sample, the positive sample The probability determination model is used to determine the probability of the target positive sample; based on the probability of the target positive sample and the probability that each sample in the sample set is the positive sample, the hybrid ranking model is trained.
  • the method of generating the positive sample may be to directly copy the positive sample.
  • the ranking score of a positive sample is 5, then 5 positive samples can be generated. For example, 5 positive samples can be copied directly, and then the previous sample set and these 5 positive samples form a new sample set.
  • a logistic regression algorithm can be used to train a positive sample probability determination model based on the new sample set, and the positive sample probability determination model is used to determine the positive sample in the new sample set The probability. Then, according to the probability of a positive sample in the new sample set and the probability of each sample in the sample set being a positive sample, a hybrid ranking model is generated. Among them, the probability that each sample in the sample set is a positive sample can be obtained through statistics or a pre-trained model.
  • the process of using the logistic regression algorithm to train a positive sample probability determination model can be as follows:
  • the model obtained at time is the positive sample probability determined model trained.
  • the positive sample probability determination model is:
  • w and b are model parameters
  • j is the feature number of each sample in the new sample set
  • the sample in the new sample set is a vector
  • the feature in the sample refers to each component in the sample.
  • the loss function is:
  • the loss function is small enough, indicating that the model converges, and the model obtained at this time is the trained model.
  • the probability of a positive sample in the new sample set can be calculated by the following formula 11):
  • N and k is the number of positive samples and the total number of samples
  • S i represents the ranking score of the i-th sample.
  • M i represents the ranking score of the i-th multimedia work in the collection of different types of multimedia work
  • Is the probability of being clicked on a multimedia work in a collection of different types of multimedia works
  • P i is the probability of being clicked on the i-th multimedia work.
  • P i can be obtained through statistics or a pre-trained model.
  • a hybrid sorting model for sorting different types of multimedia works is generated by means of a logistic regression algorithm, which is simple to implement.
  • the ranking information of each positive sample in the sample set is determined according to the degree of participation and the recommendation guidance information set by the application platform, so that different types of multimedia works can measure the ranking information based on a unified metric, which is beneficial to improve
  • the accuracy of the trained hybrid ranking model is then generated according to the ranking information of each positive sample in the sample set, and the hybrid ranking model is trained based on the new sample set.
  • the implementation method is simple.
  • Fig. 4 is a block diagram of an apparatus for recommending works according to an embodiment of the present disclosure.
  • the device includes:
  • the receiving module 41 is configured to receive a recommendation request sent by the login account of the application program, wherein the recommendation request is used to request to display a multimedia work, and the multimedia work is issued by the login account based on the associated account of the application program Multimedia works.
  • the obtaining module 42 is configured to, in response to the recommendation request received by the receiving module 41, obtain a first candidate set of works of each of the multiple types, the candidate set of first works including multimedia works published by the associated account.
  • the screening and summarizing module 43 is configured to screen each first work candidate set obtained by the obtaining module 42 and aggregate the screening results into a second work candidate set, and the second work candidate set includes the multiple types of multimedia works.
  • the sorting module 44 is configured to sort the multiple types of multimedia works in the second work candidate set summarized by the screening and summarizing module 43, and recommend the multimedia works to the login account according to the sorting result.
  • the sorting module 44 can be configured as:
  • the multimedia works in the second work candidate set are sorted.
  • the degree of participation is used to indicate the positive feedback operation or negative feedback operation performed by the account on the historical multimedia works.
  • the recommended guidance information includes At least one of recommendation information used to indicate the degree of recommendation of the historical multimedia work by the application platform or guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia work.
  • Fig. 5 is a block diagram of another device for recommending works according to an embodiment of the present disclosure.
  • the sorting module 44 may include:
  • the training sub-module 441 is configured to train a hybrid ranking model according to the participation degree and the recommendation guidance information set by the application platform, and the hybrid ranking model is used to determine the ranking sequence of the multimedia works according to the participation degree and the recommendation guidance information.
  • the sorting sub-module 442 is configured to input the multimedia works in the second candidate work set into the mixed sorting model trained by the training sub-module 441 to obtain the sorting sequence of the multimedia works in the second candidate work set.
  • the sorting module 44 may be configured to input the multiple types of multimedia works in the second work candidate set into a hybrid sorting model to obtain the sorts of the multiple types of multimedia works Sequence, the hybrid ranking model is obtained by training according to the participation degree and the recommended guidance information set by the application platform.
  • Fig. 6 is a block diagram of another work recommendation device according to an embodiment of the present disclosure. As shown in Fig. 6, based on the embodiment shown in Fig. 5, the training sub-module 441 may include:
  • the obtaining unit 4411 is configured to obtain a sample set of multiple types of multimedia works.
  • the sample set includes positive samples and negative samples.
  • the determining unit 4412 is configured to determine the ranking score of each positive sample in the sample set acquired by the acquiring unit 4411 according to the degree of participation and the recommendation guidance information set by the application platform.
  • the training unit 4413 is configured to generate a new sample set according to the ranking score and the sample set of each positive sample in the sample set determined by the determining unit 4412, and train a hybrid ranking model based on the new sample set.
  • a device for training a work recommendation hybrid ranking model is also provided.
  • the device may be as shown in FIG. 6.
  • the device includes: an acquiring unit configured to acquire multiple types of sample sets. It includes a positive sample and a negative sample.
  • the positive sample refers to a historical multimedia work that has been clicked by the account after being displayed, and the negative sample refers to a historical multimedia work that has not been clicked on by the account after being displayed;
  • the determining unit It is configured to determine the ranking score of each positive sample in the sample set according to the participation degree and the recommendation guidance information set by the application platform, and the participation degree is used to indicate the positive feedback operation performed by the account on the historical multimedia works
  • the recommendation guidance information includes at least one of recommendation information used to indicate the degree of recommendation of the application platform to historical multimedia works or guidance information used to prompt the account to perform a positive feedback operation on the historical multimedia works
  • the training unit is configured to train a hybrid ranking model according to the ranking score of each of the positive samples in the sample set and the sample set.
  • the training unit is configured to: generate a target positive sample of the same number as its ranking score for each positive sample; and train a positive sample probability determination model based on the sample set and the target positive sample
  • the positive sample probability determination model is used to determine the probability of the target positive sample; based on the probability of the target positive sample and the probability of each sample in the sample set being the positive sample, the hybrid ranking model is trained .
  • the determining unit is configured to: obtain a positive feedback operation performed by each account for each positive sample and its weight; obtain each account for each positive sample.
  • the negative feedback operation performed by the sample and its weight based on the acquired positive feedback operation and its weight, and the negative feedback operation and its weight, determine the participation of each account for each positive sample; based on all
  • the application platform sets the recommended guidance information for each of the positive samples to determine the weight of each account’s participation in each of the positive samples; The degree of participation and its weight determine the ranking score of each positive sample in the sample set.
  • the determining unit is configured to: for each feedback operation, in response to determining that the current feedback operation belongs to a low-frequency feedback operation, adjust the weight of the current feedback operation to a target high-frequency feedback operation The ratio of the occurrence frequency of the low-frequency feedback operation to the occurrence frequency of the low-frequency feedback operation.
  • the occurrence frequency of the target high-frequency feedback operation refers to the average occurrence frequency of all high-frequency feedback operations in all feedback operations currently acquired; according to each feedback operation and The adjusted weight determines the participation degree of each account for each positive sample.
  • Fig. 7 is a block diagram of a server according to an embodiment of the present disclosure.
  • the server includes a processor 710 and a memory 720 for storing instructions executable by the processor 710; wherein the processor is configured to execute the above-mentioned instructions to implement the above-mentioned work recommendation method.
  • the server usually may include other hardware according to the actual functions recommended by the work, which will not be repeated here.
  • a storage medium including instructions such as a memory 720 including instructions, which can be executed by the processor 710 to complete the above-mentioned work recommendation method.
  • the storage medium may be a non-transitory computer-readable storage medium, for example, the non-transitory computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage Equipment, etc.
  • a computer program product is further provided, which, in response to the computer program product running on an electronic device, causes the electronic device to execute the above-mentioned method for recommending works.
  • FIG. 8 is a block diagram of a device suitable for a method for recommending works according to an embodiment of the present disclosure.
  • an embodiment of the present disclosure provides a device 800 suitable for a method for recommending works, including: radio frequency (Radio Frequency (RF) circuit 810, power supply 820, processor 830, memory 840, input unit 850, display unit 860, camera 870, communication interface 880, and wireless fidelity (Wireless Fidelity, Wi-Fi) module 890 and other components.
  • RF Radio Frequency
  • the RF circuit 810 can be used to receive and send data during a communication or call. In particular, after receiving the downlink data from the base station, the RF circuit 810 sends it to the processor 830 for processing; in addition, it sends the uplink data to be sent to the base station.
  • the RF circuit 810 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier (LNA), a duplexer, and the like.
  • LNA low noise amplifier
  • the RF circuit 810 can also communicate with the network and other devices through wireless communication.
  • Wireless communication can use any communication standard or protocol, including but not limited to Global System of Mobile communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (Code Division Multiple) Access, CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), Email, Short Messaging Service (SMS), etc.
  • GSM Global System of Mobile communication
  • GPRS General Packet Radio Service
  • CDMA Code Division Multiple Access
  • WCDMA Wideband Code Division Multiple Access
  • LTE Long Term Evolution
  • Email Short Messaging Service
  • the Wi-Fi technology is a short-range wireless transmission technology.
  • the device 800 can connect to an access point (AP) through the Wi-Fi module 890, thereby achieving data network access.
  • the Wi-Fi module 890 can be used to receive and send data during the communication process.
  • the device 800 may be physically connected with other devices through the communication interface 880.
  • the communication interface 880 is connected to a communication interface of another device through a cable to realize data transmission between the device 800 and the other device.
  • the device 800 can implement communication services and send information to other contacts, the device 800 needs to have a data transmission function, that is, the device 800 needs to include a communication module inside.
  • FIG. 8 shows communication modules such as the RF circuit 810, the Wi-Fi module 890, and the communication interface 880, it is understandable that there are at least one of the above-mentioned components or other communication modules (such as Bluetooth module) for data transmission.
  • the device 800 may include an RF circuit 810 and may also include a Wi-Fi module 890; in response to the device 800 being a computer, the device 800 may include a communication interface 880 and may also include a Wi-Fi module 890; In response to the device 800 being a tablet computer, the device 800 may include a Wi-Fi module.
  • the memory 840 can be used to store software programs and modules.
  • the processor 830 executes various functional applications and data processing of the device 800 by running the software programs and modules stored in the memory 840, and after the processor 830 executes the program codes in the memory 840, the embodiments of the present disclosure can be implemented. Part or all of the process in Figure 2.
  • the memory 840 may mainly include a program storage area and a data storage area.
  • the storage program area can store operating systems, various applications (such as communication applications), and face recognition modules, etc.
  • the storage data area can store data created according to the use of the device (such as various pictures, video files and other multimedia files) , And face information templates) and so on.
  • the memory 840 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
  • a non-volatile memory such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
  • the input unit 850 may be used to receive numeric or character information input by an account, and generate key signal inputs related to account settings and function control of the device 800.
  • the input unit 850 may include a touch panel 851 and other input devices 852.
  • the touch panel 851 also called a touch screen, can collect touch operations on or near the account (for example, the account uses any suitable objects or accessories such as fingers, stylus, etc.) on the touch panel 851 or on the touch panel 851. Nearby operations), and drive the corresponding connection device according to the preset program.
  • the touch panel 851 may include two parts: a touch detection device and a touch controller.
  • the touch detection device detects the touch position of the account, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into contact coordinates, and then sends it To the processor 830, and can receive and execute the commands sent by the processor 830.
  • the touch panel 851 can be implemented in multiple types such as resistive, capacitive, infrared, and surface acoustic wave.
  • the other input device 852 may include, but is not limited to, one or more of a physical keyboard, function keys (such as volume control buttons, switch buttons, etc.), trackball, mouse, joystick, etc.
  • the display unit 860 may be used to display information input by the account or information provided to the account and various menus of the device 800.
  • the display unit 860 is the display system of the device 800, and is used to present an interface and realize human-computer interaction.
  • the display unit 860 may include a display panel 861.
  • the display panel 861 may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.
  • the touch panel 851 can cover the display panel 861. After the touch panel 851 detects a touch operation on or near it, it transmits it to the processor 830 to determine the type of the touch event, and then the processor 830 determines the type of the touch event. Corresponding visual output is provided on the display panel 861.
  • the touch panel 851 and the display panel 861 are used as two independent components to implement the input and input functions of the device 800, but in some embodiments, the touch panel 851 and the display panel 861 can be integrated And realize the input and output functions of the device 800.
  • the processor 830 is the control center of the device 800. It uses various interfaces and lines to connect various components, and executes the device 800 by running or executing software programs and/or modules stored in the memory 840 and calling data stored in the memory 840. The various functions and processing data of the device, so as to achieve a variety of equipment-based services.
  • the processor 830 may include one or more processing units. In some embodiments, the processor 830 may integrate an application processor and a modem processor, where the application processor mainly processes an operating system, an account interface, and application programs, and the modem processor mainly processes wireless communication. It can be understood that the foregoing modem processor may not be integrated into the processor 830.
  • the camera 870 is used to implement the shooting function of the device 800, and take pictures or videos.
  • the camera 870 can also be used to implement the scanning function of the device 800 and scan the scanned object (two-dimensional code/barcode).
  • the device 800 also includes a power source 820 (such as a battery) for powering various components.
  • a power source 820 such as a battery
  • the power supply 820 may be logically connected to the processor 830 through a power management system, so that functions such as charging, discharging, and power consumption can be managed through the power management system.
  • the device 800 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gates Array (FPGA), controller, microcontroller, microprocessor or other electronic components are used to perform the following operations: receive a recommendation request sent by the login account of the application program, where the recommendation request is used to request the display of multimedia works
  • the multimedia work is a multimedia work published by the login account based on the associated account of the application; in response to the recommendation request, a first candidate set of works of each of a plurality of types is obtained, and the first The work candidate set includes the multimedia works published by the associated account; each of the first work candidate sets is screened, and the screening results are aggregated into a second work candidate set, the second work candidate set including the multiple types Multimedia works; sorting the multiple types of multimedia works in the second work candidate set, and recommending multimedia works to the login account according to the sorting result.
  • ASIC application specific integrated circuits
  • the processor is specifically configured to execute the instructions to implement the following operations: sort the multimedia works in the second candidate set of works according to the degree of participation and recommended guidance information set by the application platform,
  • the degree of participation is used to indicate a positive feedback operation or a negative feedback operation performed by the account on the historical multimedia work
  • the recommendation guidance information includes recommendation information used to indicate the degree of recommendation of the application platform to the historical multimedia work or used for At least one item of guidance information that prompts the account to perform a positive feedback operation on the historical multimedia works.
  • the processor is specifically configured to execute the instructions to implement the following operations: input the multiple types of multimedia works in the second work candidate set into a mixed ranking model to obtain the multiple The sorting sequence of the type of multimedia works, the hybrid sorting model is obtained by training according to the participation degree and the recommended guidance information set by the application platform.
  • the training process of the hybrid ranking model includes: obtaining a sample set of multiple types, the sample set includes a positive sample and a negative sample, the positive sample refers to a click operation performed by the account after being displayed
  • the negative sample refers to the historical multimedia work in which the account did not perform a click operation after being displayed; each item in the sample set is determined according to the participation degree and the recommended guidance information set by the application platform
  • the ranking score of the positive sample and the hybrid ranking model is trained according to the ranking score of each positive sample in the sample set and the sample set.
  • the generating a new sample set based on the ranking information of each positive sample in the sample set and the sample set, and training a hybrid ranking model based on the new sample set includes : Generate target positive samples in the same number as its ranking score for each positive sample; train a positive sample probability determination model based on the sample set and target positive samples, and the positive sample probability determination model is used to determine the target The probability of a positive sample; based on the probability of the target positive sample and the probability of each sample in the sample set being the positive sample, the hybrid ranking model is trained.
  • the determining the ranking score of each positive sample in the sample set according to the participation degree and the recommendation guidance information set by the application platform includes: obtaining each account pair The positive feedback operation performed by each positive sample and its weight; the negative feedback operation performed by each account on each positive sample and its weight; based on the obtained positive feedback operation and its weight And negative feedback operations and their weights to determine the degree of participation of each of the accounts for each of the positive samples; based on the recommended guidance information set by the application platform for each of the positive samples, determine each of the The weight of the account’s participation in each of the positive samples; based on the account’s participation in each of the positive samples and its weight, the ranking of each of the positive samples in the sample set is determined Fraction.
  • the determining the degree of participation of each of the accounts for each of the positive samples based on the acquired positive feedback operations and their weights and the negative feedback operations and their weights includes: The feedback operation, in response to determining that the current feedback operation belongs to the low-frequency feedback operation, adjust the weight of the current feedback operation to the ratio of the occurrence frequency of the target high-frequency feedback operation to the occurrence frequency of the low-frequency feedback operation, and the target
  • the frequency of occurrence of high-frequency feedback operations refers to the average frequency of occurrence of all high-frequency feedback operations in all feedback operations currently acquired; each feedback operation and its adjusted weight are used to determine each account pair for each positive sample Participation.
  • the above-mentioned device 800 may be a server or other devices, such as a terminal device.
  • the device 800 is taken as an example of a server for description.
  • the server may include a work recommendation module and a work recommendation hybrid ranking model training module, and the work recommendation module and the work recommendation hybrid ranking model training module can call feedback mutually.
  • the work recommendation hybrid ranking model training module can train a hybrid ranking model based on samples.
  • the work recommendation module can call the hybrid ranking model trained by the work recommendation hybrid ranking model training module when it receives a recommendation request and responds to the recommendation request. Sort the multimedia works in the second work candidate set based on the hybrid sorting model to obtain the sorting result.
  • the work recommendation module can receive recommendation requests, and sort and recommend some multimedia works in response to the recommendation requests.
  • the work recommendation hybrid ranking model training module can call the work recommendation module, extract the historical processing data of the work recommendation module, and use it as a sample to train the hybrid ranking model, or, based on the historical processing data, sort the trained hybrid ranking
  • the model is further trained and optimized, and the optimized hybrid ranking model is provided to the work recommendation module to call to sort the multimedia works.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本公开关于一种作品推荐方法及服务器。其中，方法包括：接收应用程序的登录账户发送的推荐请求；响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述登录账户基于所述应用程序的关联账户发布的多媒体作品；对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。

Description

作品推荐方法及服务器
本公开要求于2020年2月20日提交的申请号为202010104322.6的发明名称为“作品推荐方法及装置、服务器和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及信息技术领域，尤其涉及一种作品推荐方法及服务器。
背景技术
随着信息技术的发展，出现了很多类型的内容(feed)，feed是将用户主动订阅的若干消息源组合在一起形成内容聚合器，帮助用户持续地获取最新的订阅源内容。
不同类型的feed混布在同一曝光场景相互竞争流量是推荐或者搜索产品的常规需求。目前，对不同类型的feed进行排序的方式是：首先对不同类型的feed分别进行排序，然后按照一定的打散规则将不同类型的feed进行混合排列。
发明内容
本公开提供一种作品推荐方法及服务器。本公开的技术方案如下：
根据本公开实施例的一方面，提供一种作品推荐方法，所述方法包括：接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
在一实施例中，所述对所述第二作品候选集中所述多个类型的多媒体作品进行排序，包括：根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进行排序，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
在一实施例中，所述根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选 集中的多媒体作品进行排序，包括：将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
根据本公开实施例的另一方面，提供一种作品推荐混合排序模型训练方法，所述方法包括：获取多个类型的样本集，所述样本集包括正样本和负样本，所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品；根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项；根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型。
在一实施例中，所述根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型，包括：为每条所述正样本生成与其排序分数等数量的目标正样本；基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
在一实施例中，所述根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，包括：获取每个所述账户对每条所述正样本执行的正向反馈操作及其权重；获取每个所述账户对每条所述正样本执行的负向反馈操作及其权重；基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度；基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重；基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数。
在一实施例中，所述基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度，包括：针对每个反馈操作，响应于确定当前反馈操作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当前获取的所有反馈操作中所有高频反馈操作的平均发生频率；根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
根据本公开实施例的另一方面，提供一种作品推荐装置，所述装置包括：接收模块，被配置为接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体 作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；获取模块，被配置为响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；筛选汇总模块，被配置为对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；排序模块，被配置为对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
在一实施例中，所述排序模块，被配置为：根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进行排序，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
在一实施例中，所述排序模块包括：将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
在一实施例中，所述排序模块包括：训练子模块，被配置为根据所述参与度和所述应用平台设置的所述推荐引导信息，训练出混合排序模型，所述混合排序模型用于根据所述参与度和所述推荐引导信息确定出多媒体作品的排序序列；排序子模块，被配置为将所述第二作品候选集中所述多个类型的多媒体作品输入所述训练子模块训练出的所述混合排序模型，得到所述第二作品候选集中所述多个类型的多媒体作品的排序序列。
在一实施例中，所述训练子模块包括：获取单元，被配置为获取多个类型的样本集，所述样本集包括正样本和负样本，所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品；确定单元，被配置为根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项；训练单元，被配置为根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型。
在一实施例中，所述训练单元，被配置为：为每条所述正样本生成与其排序分数等数量的目标正样本；基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
在一实施例中，所述确定单元，被配置为：获取每个所述账户对每条所述正样本执行的正向反馈操作及其权重；获取每个所述账户对每条所述正样本执行的负向反馈操作及其权重；基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度；基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重；基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数。
在一实施例中，所述确定单元，被配置为：针对每个反馈操作，响应于确定当前反馈操作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当前获取的所有反馈操作中所有高频反馈操作的平均发生频率；根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
根据本公开实施例的另一方面，提供一种作品推荐混合排序模型训练装置，所述装置包括：获取单元，被配置为获取多个类型的样本集，所述样本集包括正样本和负样本，所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品；确定单元，被配置为根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项；训练单元，被配置为根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型。
在一实施例中，所述训练单元，被配置为：为每条所述正样本生成与其排序分数等数量的目标正样本；基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
在一实施例中，所述确定单元，被配置为：获取每个所述账户对每条所述正样本执行的正向反馈操作及其权重；获取每个所述账户对每条所述正样本执行的负向反馈操作及其权重；基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度；基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重；基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数。
在一实施例中，所述确定单元，被配置为：针对每个反馈操作，响应于确定当前反馈操 作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当前获取的所有反馈操作中所有高频反馈操作的平均发生频率；根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
根据本公开实施例的另一方面，提供一种服务器，包括：
处理器；
用于存储所述处理器可执行指令的存储器；
其中，所述处理器被配置为执行所述指令，以实现如下操作：接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
根据本公开实施例的另一方面，提供一种存储介质，响应于所述存储介质中的指令由服务器的处理器执行，使得服务器能够执行如下操作：接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
通过接收应用程序的登录账户发送的推荐请求，并响应于该推荐请求，获取关联账户发布的多个类型的第一作品候选集，然后在至少基于服务器处理参数分别对各第一作品候选集进行筛选之后，将筛选结果汇总为第二作品候选集，最后对第二作品候选集中的多媒体作品进行排序，并根据排序结果，向客户端推荐多媒体作品，由于该实施例是对包含多个类型的第一作品候选集进行筛选之后汇总得到的第二作品候选集中的作品进行排序，即该实施例是对多个类型的作品统一进行排序，有利于提高排序结果的准确性，从而提高根据排序结果推荐多媒体作品的准确性。
附图说明
图1是本公开一实施例示出的一种作品推荐方法的流程图。
图2是本公开一实施例示出的一种训练混合排序模型的流程图。
图3是本公开一实施例示出的一种确定样本集中每条正样本的排序分数的流程图。
图4是本公开一实施例示出的一种作品推荐装置的框图。
图5是本公开一实施例示出的另一种作品推荐装置的框图。
图6是本公开一实施例示出的另一种作品推荐装置的框图。
图7是本公开一实施例示出的一种服务器的框图。
图8是本公开一实施例示出的一种适用于作品推荐方法的设备的框图。
具体实施方式
本公开的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。
图1是本公开一实施例示出的一种作品推荐方法的流程图，如图1所示，该作品推荐方法包括以下内容：
在S101中，接收应用程序的登录账户发送的推荐请求，其中，该推荐请求用于请求在目标页面上展示多媒体作品，目标页面用于展示通过应用程序与登录账户建立有社交关系的关联账户所发布的多媒体作品。
所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品。所述登录账户与关联账户基于所述应用程序建立有社交关系。
其中，目标页面可以包括但不局限于关注页，还可以包括同城页等。多媒体作品可以包括但不局限于直播、短视频等作品。
当用户登录应用程序时，该应用程序可以向服务器发送该登录账户的推荐请求，该应用程序可以包括但不局限于发布多媒体作品的APP。
在本公开的实施例中，响应于目标页面是关注页，与登录账户建立有社交关系的关联账户可以包括登录账户的关注账户。响应于目标页面是同城页，与登录账户建立了有社交关系的关联账户可以包括与登录账户位于同一个城市的账户。
在S102中，响应于该推荐请求，从作品库中获取属于关联账户发布多媒体作品的各个类型的第一作品候选集。
该S102为响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集的一种可能实施方式，所述第一作品候选集包括所述关联账户发布的多媒体作品。该各个类型是指多个类型中每个类型。
服务器在接收该推荐请求之后，响应于该推荐请求，可以从作品库中获取第一作品候选 集，其中，各个类型的第一作品候选集可以包括但不局限于直播类第一作品候选集、短视频类第一作品候选集。该第一作品候选集也可以位于其他地址，例如，第一作品候选集位于服务器本地存储中，本公开实施例对此不作具体限定。
在S103中，在至少基于服务器处理参数分别对各第一作品候选集进行筛选之后，将筛选结果汇总为第二作品候选集。
该S103为对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集的一种可能实施方式，所述第二作品候选集包括所述多个类型的多媒体作品。在筛选时，能够基于服务器处理参数，也可以基于其他筛选规则，本公开实施例对此不作具体限定。
在获取各个类型的第一作品候选集之后，可以基于服务器的性能例如服务器处理参数，分别对各第一作品候选集进行筛选，并将筛选结果汇总为第二作品候选集。
在S104中，对第二作品候选集中的多媒体作品进行排序，并根据排序结果，向客户端推荐多媒体作品。
该S104为对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品的过程。该过程中向登录账号推荐多媒体作品，则可以向登录有该登录账号的客户端发送多媒体作品。该推荐过程可以为按照排序结果，在客户端上显示多媒体作品。例如，排序结果靠前的多媒体作品优先显示。
在本公开的实施例中，可以根据参与度和应用平台设置的推荐引导信息，对第二作品候选集中的多媒体作品进行排序，即采用统一的标准对第二作品候选集中的多媒体作品进行排序，有利于提高排序结果的准确性。
其中，参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作。正向反馈操作可以包括但不局限于观看操作、点赞操作、关注操作、评论操作等，负向反馈操作可以包括但不局限于包括举报操作等。
其中，推荐引导信息可以包括用于表示应用平台对历史多媒体作品的推荐度的推荐信息和用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中的至少一项。也即是包括三种情况：推荐引导信息可以包括用于表示应用平台对历史多媒体作品的推荐度的推荐信息。推荐引导信息可以包括用于提示账户对历史多媒体作品执行正向反馈操作的引导信息。推荐引导信息可以包括用于表示应用平台对历史多媒体作品的推荐度的推荐信息和用于提示账户对历史多媒体作品执行正向反馈操作的引导信息。
在本公开的实施例中，根据参与度和应用平台设置的推荐引导信息，对第二作品候选集中的多媒体作品进行排序，可以包括：将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所 述参与度和所述应用平台设置的所述推荐引导信息训练得到。
其中，该训练过程可以为：根据参与度和应用平台设置的推荐引导信息，训练出混合排序模型，混合排序模型用于根据参与度和推荐引导信息确定出多媒体作品的排序序列。该训练过程可以预先完成，在对第二作品候选集进行排序时直接使用训练好的混合排序模型。该训练过程也可以在需要对第二作品候选集进行排序时执行。
本公开的实施例，通过接收应用程序的登录账户发送的推荐请求，并响应于该推荐请求，从作品库中获取属于关联账户发布多媒体作品的各个类型的第一作品候选集，然后在至少基于服务器处理参数分别对各第一作品候选集进行筛选之后，将筛选结果汇总为第二作品候选集，最后对第二作品候选集中的多媒体作品进行排序，并根据排序结果，向客户端推荐多媒体作品，由于本公开的实施例是对包含多个类型的第一作品候选集进行筛选之后汇总得到的第二作品候选集中的作品进行排序，即本公开的实施例是对各个类型的作品统一进行排序，有利于提高排序结果的准确性，从而提高根据排序结果推荐多媒体作品的准确性。
为了使用混合排序模型对第二作品候选集中的多媒体作品进行排序，本公开的实施例需要预先训练出混合排序模型，图2是本公开一实施例示出的一种训练混合排序模型的流程图，如图2所示，该训练混合排序模型的过程可以包括以下内容：
在S201中，获取多个类型多媒体作品的样本集，该样本集包括正样本和负样本，正样本是指在目标页面中展现后被账户执行点击操作的历史多媒体作品，负样本是指在目标页面中展现后账户未执行点击操作的历史多媒体作品。
该S201为获取多个类型的样本集的过程，样本集中包括样本，每个样本可以为多个类型中任一类型的多媒体作品。所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品。该展现过程为展现在目标页面，也可能为展现在其他页面，本公开的实施例对此不作限定。
其中，多个类型多媒体作品可以包括但不局限于直播、短视频等作品。
在本公开的实施例中，上述展现日志包括用户标识(userId)和作品(Item)，响应于某个账户的Item被展现(show)，则该账户的Item就是一条样本。响应于该Item被账户执行点击操作，则该Item就是一条正样本。响应于该Item未被账户执行点击操作，则该Item就是一条负样本。
在本公开的实施例中，可以获取多个类型多媒体作品的展现日志，响应于根据展现日志确定对应的多媒体作品展现后被账户执行点击操作，则生成一条正样本，且标签(label)记为1。响应于根据展现日志确定对应的多媒体作品展现后该账户未执行点击操作，则生成一条负样本，且标签(label)记为0。
在S202中，根据参与度和应用平台设置的推荐引导信息，确定样本集中每条正样本的排序分数。
其中，账户的参与度可以基于账户对历史多媒体作品执行的正向反馈操作及其权重和负向反馈操作及其权重来确定，应用平台设置的推荐引导信息可以由应用平台基于生态因素设置，例如生态因素可以包括但不局限于流量普惠程度等，也可以基于其他因素设置。
例如，如图3所示，确定样本集中每条正样本的排序分数可以包括：
在S2021中，获取每个账户对每条正样本的执行的正向反馈操作及其权重和每个账户对每条正样本的执行的负向反馈操作及其权重，并根据每个账户对每条正样本的执行的正向反馈操作及其权重和每个账户对每条正样本的执行的负向反馈操作及其权重确定每个账户对每条正样本的参与度。
该S2021中确定参与度的过程为基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度的过程。
其中，正向反馈操作可以包括但不局限于观看操作、点赞操作、关注操作、评论操作等，负向反馈操作可以包括但不局限于包括举报操作等。
在本公开的实施例中，可以通过留存归因算法确定账户对每条正样本的执行的负向反馈操作的权重或负向反馈操作的权重，然后根据每个反馈操作及其权重进行加权运算，得到每个账户对每条正样本的参与度。
在一些实施例中，该行为为一种反馈操作，可以为正向反馈操作，也可以为负向反馈操作。针对每个反馈操作，响应于确定当前反馈操作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当前获取的所有反馈操作中所有高频反馈操作的平均发生频率；根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
也即是，针对每个行为，响应于确定当前行为属于低频行为，则将当前行为的权重调整为目标高频行为的发生频率与低频行为的发生频率的比值，并根据每个行为及其调整后的权重确定每个账户对每条正样本的参与度。其中，目标高频行为的发生频率是指当前获取的所有行为中所有高频行为的平均发生频率。
例如，当前获取的所有行为是点赞和评论，其中，点赞为高频行为，评论为低频行为，预先统计的点赞的发生频率为0.1，预先统计的评论的发生频率为0.001，则可以将评论的权重调整为100。需要说明的是，该举例中涉及的行为及数值仅为一个示例，在实际应用中，可以根据需要进行调整。
在S2022中，使用应用平台为每条正样本设置的推荐引导信息来确定每个账户对每条正样本的参与度的权重。
该S2022为基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重的过程。
其中，账户对不同类型多媒体作品的接受程度不一样，例如，账户对短视频的接受程度高于对直播的接受程度，这时可以增大账户对短视频的参与度的权重。
在S2023中，根据每个账户对每条正样本的参与度及其权重，得到样本集中每条正样本的排序分数。
该S2023为基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数的过程。
在确定每个账户对每条正样本的参与度的权重之后，可以根据每个账户对每条正样本的参与度及其权重，得到样本集中每条正样本的排序分数。
在本公开的实施例中，根据参与度和应用平台设置的推荐引导信息来确定样本集中每条正样本的排序分数，使得不同类型的多媒体作品可以基于统一的标准来度量排序信息，有利于提高训练出的混合排序模型的准确性。
另外，由于参与度和平台设置的推荐度可以包括多个维度的信息，这些信息可以很好地描述应用场景的特征，从而可以进一步提高训练出的混合排序模型的准确性，也即进一步提高通过该混合排序模型确定的排序序列的准确性。
在S203中，根据样本集中每条正样本的排序分数和样本集生成新的样本集，并基于新的样本集训练出混合排序模型。
该S203为根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出所述混合排序模型的过程。该过程可以不生成新的样本集，直接使用样本集和生成的正样本，该生成的正样本可以称为目标正样本。该混合排序模型的训练过程可以包括：为每条所述正样本生成与其排序分数等数量的目标正样本；基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
在本公开的实施例中，可以为每条正样本生成与其排序分数等数量的正样本，并根据样本集和生成的正样本获得新的样本集。
其中，生成正样本的方式可以为直接复制正样本。
例如，某条正样本的排序分数为5，则可以生成5条正样本，例如，直接复制出5条该正样本，然后由之前的样本集和这5条正样本组成新的样本集。
在本公开的实施例中，在获得新的样本集之后，可以基于新的样本集，采用逻辑回归算法训练出正样本概率确定模型，该正样本概率确定模型用于确定新的样本集中正样本的概率。然后，根据新的样本集中正样本的概率和样本集中每条样本为正样本的概率，生成混合排序模型。其中，样本集中每条样本为正样本的概率可以通过统计获得，也可以通过预先训练的模型获得。
其中，采用逻辑回归算法训练出正样本概率确定模型的过程可以为：
将新的样本集中的样本输入正样本概率确定模型，得到正样本的概率，根据得到的正样本概率计算损失函数，利用该损失函数更新正样本概率确定模型的参数，直至损失函数足够小，此时得到的模型为训练出的正样本概率确定模型。
在本公开的实施例中，正样本概率确定模型为：
Figure PCTCN2021076195-appb-000001
其中，w和b为模型参数，j为新的样本集中每个样本中的特征编号，新的样本集中的样本为向量，样本中的特征指样本中的每个分量。
在本公开的实施例中，损失函数为：
Figure PCTCN2021076195-appb-000002
损失函数足够小，说明模型收敛，此时得到的模型即为训练出的模型。
在本公开的实施例中，在新的样本集中正样本的概率可以通过以下公式11)计算：
Figure PCTCN2021076195-appb-000003
其中，N和k分别为样本的总个数和正样本的个数，S i表示第i个样本的排序分数。
而由逻辑回归的性质可得以下公式12)：
Figure PCTCN2021076195-appb-000004
根据上述公式11)和公式12)可得：
Figure PCTCN2021076195-appb-000005
进而可以得到混合排序模型为：
Figure PCTCN2021076195-appb-000006
其中，M i表示不同类型多媒体作品集合中第i个多媒体作品的排序分数，
Figure PCTCN2021076195-appb-000007
为不同类型多媒体作品集合中多媒体作品被点击的概率，P i为第i个多媒体作品被点击的概率，P i可以通过统计获得，也可以通过预先训练出的模型来获得。
在本公开的实施例中，借助逻辑回归算法生成对不同类型多媒体作品进行排序的混合排序模型，实现方式简单。
本公开的实施例，通过根据参与度和应用平台设置的推荐引导信息来确定样本集中每条正样本的排序信息，使得不同类型的多媒体作品可以基于统一的度量标准来度量排序信息，有利于提高训练出的混合排序模型的准确度，然后根据样本集中每条正样本的排序信息生成新的样本集，并基于新的样本集训练出混合排序模型，实现方式简单。
图4是本公开一实施例示出的一种作品推荐装置的框图。参照图4，该装置包括：
接收模块41被配置为接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品。
获取模块42被配置为响应于接收模块41接收的推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品。
筛选汇总模块43被配置为对获取模块42获取的各第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品。
排序模块44被配置为对筛选汇总模块43汇总后的第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
其中，排序模块44可以被配置为：
根据参与度和应用平台设置的推荐引导信息，对第二作品候选集中的多媒体作品进行排序，参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，推荐引导信息包括用于表示应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
图5是本公开一实施例示出的另一种作品推荐装置的框图，如图5所示，在图4所示实施例的基础上，排序模块44可以包括：
训练子模块441被配置为根据参与度和应用平台设置的推荐引导信息，训练出混合排序模型，混合排序模型用于根据参与度和推荐引导信息确定出多媒体作品的排序序列。
排序子模块442被配置为将第二作品候选集中的多媒体作品输入训练子模块441训练出的混合排序模型，得到第二作品候选集中多媒体作品的排序序列。
在图4所示实施例的基础上，排序模块44可以被配置为将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
图6是本公开一实施例示出的另一种作品推荐装置的框图，如图6所示，在图5所示实施例的基础上，训练子模块441可以包括：
获取单元4411被配置为获取多个类型多媒体作品的样本集，样本集包括正样本和负样 本，正样本是指在目标页面中展现后被账户执行点击操作的历史多媒体作品，负样本是指在目标页面中展现后账户未执行点击操作的历史多媒体作品。
确定单元4412被配置为根据参与度和应用平台设置的推荐引导信息，确定获取单元4411获取的样本集中每条正样本的排序分数。
训练单元4413被配置为根据确定单元4412确定的样本集中每条正样本的排序分数和样本集生成新的样本集，并基于新的样本集训练出混合排序模型。
在一些实施例中，还提供了一种作品推荐混合排序模型训练装置，该装置可以如图6所示，该装置包括：获取单元，被配置为获取多个类型的样本集，所述样本集包括正样本和负样本，所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品；确定单元，被配置为根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项；训练单元，被配置为根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型。
在一实施例中，所述训练单元，被配置为：为每条所述正样本生成与其排序分数等数量的目标正样本；基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
在一实施例中，所述确定单元，被配置为：获取每个所述账户对每条所述正样本执行的正向反馈操作及其权重；获取每个所述账户对每条所述正样本执行的负向反馈操作及其权重；基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度；基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重；基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数。
在一实施例中，所述确定单元，被配置为：针对每个反馈操作，响应于确定当前反馈操作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当前获取的所有反馈操作中所有高频反馈操作的平均发生频率；根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
图7是本公开一实施例示出的一种服务器的框图。如图7所示，该服务器包括处理器710、用于存储处理器710可执行指令的存储器720；其中，处理器被配置为执行上述指令，以实现上述作品推荐方法。除了图7所示的处理器710及存储器720之外，该服务器通常根据作品推荐的实际功能，还可以包括其他硬件，对此不再赘述。
在一实施例中，还提供了一种包括指令的存储介质，例如包括指令的存储器720，上述指令可由处理器710执行以完成上述作品推荐方法。在一些实施例中，存储介质可以是非临时性计算机可读存储介质，例如，非临时性计算机可读存储介质可以是ROM、随机存取存储器(RAM)、CD-ROM、磁带、软盘和光数据存储设备等。
在一实施例中，还提供一种计算机程序产品，响应于计算机程序产品在电子设备上运行，使得电子设备执行上述作品推荐方法。
图8是本公开一实施例示出的一种适用于作品推荐方法的设备的框图，如图8所示，本公开实施例给出一种适用于作品推荐方法的设备800，包括：射频(Radio Frequency，RF)电路810、电源820、处理器830、存储器840、输入单元850、显示单元860、摄像头870、通信接口880、以及无线保真(Wireless Fidelity，Wi-Fi)模块890等部件。本领域技术人员可以理解，图8中示出的设备的结构并不构成对设备的限定，本公开实施例提供的设备可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件布置。
下面结合图8对设备800的各个构成部件进行具体的介绍：
RF电路810可用于通信或通话过程中，数据的接收和发送。特别地，RF电路810在接收到基站的下行数据后，发送给处理器830处理；另外，将待发送的上行数据发送给基站。通常，RF电路810包括但不限于天线、至少一个放大器、收发信机、耦合器、低噪声放大器(Low Noise Amplifier，LNA)、双工器等。
此外，RF电路810还可以通过无线通信与网络和其他设备通信。无线通信可以使用任一通信标准或协议，包括但不限于全球移动通讯系统(Global System of Mobile communication，GSM)、通用分组无线服务(General Packet Radio Service，GPRS)、码分多址(Code Division Multiple Access，CDMA)、宽带码分多址(Wideband Code Division Multiple Access，WCDMA)、长期演进(Long Term Evolution，LTE)、电子邮件、短消息服务(Short Messaging Service，SMS)等。
Wi-Fi技术属于短距离无线传输技术，设备800通过Wi-Fi模块890可以连接接入点(Access Point，AP)，从而实现数据网络的访问。Wi-Fi模块890可用于通信过程中，数据的 接收和发送。
设备800可以通过通信接口880与其他设备实现物理连接。在一些实施例中，通信接口880与其他设备的通信接口通过电缆连接，实现设备800和其他设备之间的数据传输。
由于在本公开实施例中，设备800能够实现通信业务，向其他联系人发送信息，因此设备800需要具有数据传输功能，即设备800内部需要包含通信模块。虽然图8示出了RF电路810、Wi-Fi模块890、和通信接口880等通信模块，但是可以理解的是，设备800中存在上述部件中的至少一个或者其他用于实现通信的通信模块(如蓝牙模块)，以进行数据传输。
例如，响应于设备800为手机，设备800可以包含RF电路810，还可以包含Wi-Fi模块890；响应于设备800为计算机，设备800可以包含通信接口880，还可以包含Wi-Fi模块890；响应于设备800为平板电脑，设备800可以包含Wi-Fi模块。
存储器840可用于存储软件程序以及模块。处理器830通过运行存储在存储器840的软件程序以及模块，从而执行设备800的各种功能应用以及数据处理，并且处理器830执行存储器840中的程序代码后，可以实现本公开实施例图1、图2中的部分或全部过程。
在一些实施例中，存储器840可以主要包括存储程序区和存储数据区。其中，存储程序区可存储操作系统、各种应用程序(比如通信应用)以及人脸识别模块等；存储数据区可存储根据设备的使用所创建的数据(比如各种图片、视频文件等多媒体文件，以及人脸信息模板)等。
此外，存储器840可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他易失性固态存储器件。
输入单元850可用于接收账户输入的数字或字符信息，以及产生与设备800的账户设置以及功能控制有关的键信号输入。
在一些实施例中，输入单元850可包括触控面板851以及其他输入设备852。
其中，触控面板851，也称为触摸屏，可收集账户在其上或附近的触摸操作(比如账户使用手指、触笔等任何适合的物体或附件在触控面板851上或在触控面板851附近的操作)，并根据预先设定的程式驱动相应的连接装置。在一些实施例中，触控面板851可以包括触摸检测装置和触摸控制器两个部分。其中，触摸检测装置检测账户的触摸方位，并检测触摸操作带来的信号，将信号传送给触摸控制器；触摸控制器从触摸检测装置上接收触摸信息，并将它转换成触点坐标，再送给处理器830，并能接收处理器830发来的命令并加以执行。此外，可以采用电阻式、电容式、红外线以及表面声波等多个类型实现触控面板851。
在一些实施例中，其他输入设备852可以包括但不限于物理键盘、功能键(比如音量控制按键、开关按键等)、轨迹球、鼠标、操作杆等中的一种或多种。
显示单元860可用于显示由账户输入的信息或提供给账户的信息以及设备800的各种菜单。显示单元860即为设备800的显示系统，用于呈现界面，实现人机交互。
显示单元860可以包括显示面板861。在一些实施例中，显示面板861可以采用液晶显示屏(Liquid Crystal Display，LCD)、有机发光二极管(Organic Light-Emitting Diode，OLED)等形式来配置。
进一步的，触控面板851可覆盖显示面板861，触控面板851检测到在其上或附近的触摸操作后，传送给处理器830以确定触摸事件的类型，随后处理器830根据触摸事件的类型在显示面板861上提供相应的视觉输出。
虽然在图8中，触控面板851与显示面板861是作为两个独立的部件来实现设备800的输入和输入功能，但是在某些实施例中，可以将触控面板851与显示面板861集成而实现设备800的输入和输出功能。
处理器830是设备800的控制中心，利用各种接口和线路连接各个部件，通过运行或执行存储在存储器840内的软件程序和/或模块，以及调用存储在存储器840内的数据，执行设备800的各种功能和处理数据，从而实现基于设备的多种业务。
在一些实施例中，处理器830可包括一个或多个处理单元。在一些实施例中，处理器830可集成应用处理器和调制解调处理器，其中，应用处理器主要处理操作系统、账户界面和应用程序等，调制解调处理器主要处理无线通信。可以理解的是，上述调制解调处理器也可以不集成到处理器830中。
摄像头870，用于实现设备800的拍摄功能，拍摄图片或视频。摄像头870还可以用于实现设备800的扫描功能，对扫描对象(二维码/条形码)进行扫描。
设备800还包括用于给各个部件供电的电源820(比如电池)。在一些实施例中，电源820可以通过电源管理系统与处理器830逻辑相连，从而通过电源管理系统实现管理充电、放电、以及功耗等功能。
在一实施例中，设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行如下操作：接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；对所述第二作品候选集中所述多个类型的多 媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
在一些实施例中，所述处理器具体被配置为执行所述指令，以实现如下操作：根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进行排序，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
在一些实施例中，所述处理器具体被配置为执行所述指令，以实现如下操作：将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
在一些实施例中，所述混合排序模型的训练过程包括：获取多个类型的样本集，所述样本集包括正样本和负样本，所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品；根据所述参与度和所述应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数；根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出所述混合排序模型。
在一些实施例中，所述根据所述样本集中每条所述正样本的所述排序信息和所述样本集生成新的样本集，并基于所述新的样本集训练出混合排序模型，包括：为每条所述正样本生成与其排序分数等数量的目标正样本；基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
在一些实施例中，所述根据所述参与度和所述应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，包括：获取每个所述账户对每条所述正样本执行的正向反馈操作及其权重；获取每个所述账户对每条所述正样本执行的负向反馈操作及其权重；基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度；基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重；基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数。
在一些实施例中，所述基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度，包括：针对每个反馈操作，响应于确定当前反馈操作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当 前获取的所有反馈操作中所有高频反馈操作的平均发生频率；根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
上述设备800可以为服务器，也可以为其他设备，比如为终端设备。在一些实施例中，以该设备800为服务器为例进行说明，该服务器可以包括作品推荐模块和作品推荐混合排序模型训练模块，该作品推荐模块和作品推荐混合排序模型训练模块能够相互调用反馈。例如，作品推荐混合排序模型训练模块能够基于样本训练出混合排序模型，作品推荐模块在接收到推荐请求，响应于该推荐请求时，能够调用作品推荐混合排序模型训练模块训练出的混合排序模型，基于该混合排序模型对第二作品候选集中的多媒体作品进行排序得到排序结果。又例如，该作品推荐模块能够接收推荐请求，并响应于推荐请求对一些多媒体作品进行排序和推荐。该作品推荐混合排序模型训练模块能够调用该作品推荐模块，提取该作品推荐模块的历史处理数据，将其作为样本来训练混合排序模型，或者，基于该历史处理数据，对已训练好的混合排序模型进行进一步训练优化，将优化好的混合排序模型提供给作品推荐模块调用以对多媒体作品进行排序。

Claims (14)

  1. 一种作品推荐方法，其中，所述方法包括：
    接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；
    响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；
    对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；
    对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
  2. 根据权利要求1所述的作品推荐方法，其中，所述对所述第二作品候选集中所述多个类型的多媒体作品进行排序，包括：
    根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进行排序，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
  3. 根据权利要求2所述的作品推荐方法，其中，所述根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进行排序，包括：
    将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
  4. 一种作品推荐混合排序模型训练方法，其中，所述方法包括：
    获取多个类型的样本集，所述样本集包括正样本和负样本，所述正样本是指展现后被所述账户执行点击操作的历史多媒体作品，所述负样本是指展现后所述账户未执行点击操作的历史多媒体作品；
    根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项；
    根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型。
  5. 根据权利要求4所述的作品推荐混合排序模型训练方法，其中，所述根据所述样本集中每条所述正样本的所述排序分数和所述样本集，训练出混合排序模型，包括：
    为每条所述正样本生成与其排序分数等数量的目标正样本；
    基于所述样本集和目标正样本，训练出正样本概率确定模型，所述正样本概率确定模型用于确定所述目标正样本的概率；
    基于所述目标正样本的概率和所述样本集中每条样本为所述正样本的概率，训练出所述混合排序模型。
  6. 根据权利要求4所述的作品推荐混合排序模型训练方法，其中，所述根据参与度和应用平台设置的所述推荐引导信息，确定所述样本集中每条所述正样本的排序分数，包括：
    获取每个所述账户对每条所述正样本执行的正向反馈操作及其权重；
    获取每个所述账户对每条所述正样本执行的负向反馈操作及其权重；
    基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度；
    基于所述应用平台为每条所述正样本设置的所述推荐引导信息，确定每个所述账户对每条所述正样本的参与度的权重；
    基于每个所述账户对每条所述正样本的参与度及其权重，确定所述样本集中每条所述正样本的所述排序分数。
  7. 根据权利要求6所述的作品推荐混合排序模型训练方法，其中，所述基于获取到的正向反馈操作及其权重和负向反馈操作及其权重，确定每个所述账户对每条所述正样本的参与度，包括：
    针对每个反馈操作，响应于当前反馈操作属于低频反馈操作，则将所述当前反馈操作的所述权重调整为目标高频反馈操作的发生频率与所述低频反馈操作的发生频率的比值，所述目标高频反馈操作的发生频率是指当前获取的所有反馈操作中所有高频反馈操作的平均发生频率；
    根据每个反馈操作及其调整后的权重确定每个所述账户对每条所述正样本的参与度。
  8. 一种作品推荐装置，其中，所述装置包括：
    接收模块，被配置为接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；
    获取模块，被配置为响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；
    筛选汇总模块，被配置为对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；
    排序模块，被配置为对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
  9. 根据权利要求8所述的作品推荐装置，其中，所述排序模块，被配置为：
    根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进行排序，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
  10. 根据权利要求9所述的作品推荐装置，其中，所述排序模块，被配置为将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
  11. 一种服务器，其中，包括：
    处理器；
    用于存储所述处理器可执行指令的存储器；
    其中，所述处理器被配置为执行所述指令，以实现如下操作：
    接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；
    响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；
    对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；
    对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
  12. 根据权利要求11所述的服务器，其中，所述处理器具体被配置为执行所述指令，以实现如下操作：
    根据参与度和应用平台设置的推荐引导信息，对所述第二作品候选集中的多媒体作品进 行排序，所述参与度用于表示账户对历史多媒体作品执行的正向反馈操作或负向反馈操作，所述推荐引导信息包括用于表示所述应用平台对历史多媒体作品的推荐度的推荐信息或用于提示账户对历史多媒体作品执行正向反馈操作的引导信息中至少一项。
  13. 根据权利要求12所述的服务器，其中，所述处理器具体被配置为执行所述指令，以实现如下操作：
    将所述第二作品候选集中所述多个类型的多媒体作品输入混合排序模型，得到所述多个类型的多媒体作品的排序序列，所述混合排序模型根据所述参与度和所述应用平台设置的所述推荐引导信息训练得到。
  14. 一种存储介质，其中，响应于所述存储介质中的指令由服务器的处理器执行，使得服务器能够执行如下操作：
    接收应用程序的登录账户发送的推荐请求，其中，所述推荐请求用于请求展示多媒体作品，所述多媒体作品为所述登录账户基于所述应用程序的关联账户所发布的多媒体作品；
    响应于所述推荐请求，获取多个类型中每个类型的第一作品候选集，所述第一作品候选集包括所述关联账户发布的多媒体作品；
    对各所述第一作品候选集进行筛选，将筛选结果汇总为第二作品候选集，所述第二作品候选集包括所述多个类型的多媒体作品；
    对所述第二作品候选集中所述多个类型的多媒体作品进行排序，并根据排序结果，向所述登录账号推荐多媒体作品。
PCT/CN2021/076195 2020-02-20 2021-02-09 作品推荐方法及服务器 WO2021164643A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2022547740A JP2023512692A (ja) 2020-02-20 2021-02-09 作品の推奨方法及びサーバー
US17/891,465 US20220398277A1 (en) 2020-02-20 2022-08-19 Method for recommending works and server

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010104322.6A CN111339327B (zh) 2020-02-20 2020-02-20 作品推荐方法及装置、服务器和存储介质
CN202010104322.6 2020-02-20

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/891,465 Continuation US20220398277A1 (en) 2020-02-20 2022-08-19 Method for recommending works and server

Publications (1)

Publication Number Publication Date
WO2021164643A1 true WO2021164643A1 (zh) 2021-08-26

Family

ID=71183554

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/076195 WO2021164643A1 (zh) 2020-02-20 2021-02-09 作品推荐方法及服务器

Country Status (4)

Country Link
US (1) US20220398277A1 (zh)
JP (1) JP2023512692A (zh)
CN (1) CN111339327B (zh)
WO (1) WO2021164643A1 (zh)

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111339327B (zh) * 2020-02-20 2024-05-07 北京达佳互联信息技术有限公司 作品推荐方法及装置、服务器和存储介质
CN111737579B (zh) * 2020-06-28 2024-06-25 北京达佳互联信息技术有限公司 对象推荐方法、装置及电子设备、存储介质
CN113886674B (zh) * 2020-07-01 2024-09-06 北京达佳互联信息技术有限公司 资源推荐方法、装置、电子设备及存储介质
CN113961792B (zh) * 2020-07-15 2024-07-09 北京达佳互联信息技术有限公司 一种资源推荐系统、方法、电子设备及存储介质
CN111859129B (zh) * 2020-07-20 2024-05-10 游艺星际(北京)科技有限公司 对象推荐方法、装置、电子设备和存储介质
CN111859141B (zh) * 2020-07-28 2023-09-26 北京达佳互联信息技术有限公司 内容推送方法、装置、服务器及存储介质
CN112000821B (zh) * 2020-08-21 2024-03-26 北京达佳互联信息技术有限公司 多媒体信息推送方法、装置、服务器及存储介质
CN112115354B (zh) * 2020-09-02 2024-09-20 北京达佳互联信息技术有限公司 信息处理方法、装置、服务器及存储介质
CN112084150B (zh) * 2020-09-09 2024-07-26 北京百度网讯科技有限公司 模型训练、数据检索方法，装置，设备以及存储介质
CN112272302A (zh) * 2020-09-17 2021-01-26 北京达佳互联信息技术有限公司 多媒体资源的展示方法、装置、系统及存储介质
CN112199586A (zh) * 2020-09-29 2021-01-08 北京字节跳动网络技术有限公司 一种作品分发方法、装置、计算机设备及存储介质
CN114329048A (zh) * 2020-09-30 2022-04-12 北京达佳互联信息技术有限公司 素材确定方法、装置及电子设备
CN112269917B (zh) * 2020-10-09 2023-10-31 北京达佳互联信息技术有限公司 一种媒体资源展示方法及装置、设备、系统、存储介质
CN117857884B (zh) * 2024-03-07 2024-06-07 北京字跳网络技术有限公司 多媒体内容的处理方法、装置及终端设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150319181A1 (en) * 2014-04-30 2015-11-05 Twitter, Inc. Application Graph Builder
CN107609146A (zh) * 2017-09-21 2018-01-19 北京达佳互联信息技术有限公司 信息展示方法、装置、终端及服务器
CN108334536A (zh) * 2017-11-30 2018-07-27 中国电子科技集团公司电子科学研究院 一种信息推荐方法、设备和存储介质
CN109902708A (zh) * 2018-12-29 2019-06-18 华为技术有限公司 一种推荐模型训练方法及相关装置
CN111339327A (zh) * 2020-02-20 2020-06-26 北京达佳互联信息技术有限公司 作品推荐方法及装置、服务器和存储介质

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9230212B2 (en) * 2012-02-02 2016-01-05 Peel Technologies, Inc. Content based recommendation system
US8996530B2 (en) * 2012-04-27 2015-03-31 Yahoo! Inc. User modeling for personalized generalized content recommendations
CN108460060A (zh) * 2017-02-22 2018-08-28 合网络技术（北京）有限公司 视频推荐方法及装置
US20180336645A1 (en) * 2017-05-22 2018-11-22 Google Inc. Using machine learning to recommend live-stream content
US11954161B2 (en) * 2018-05-31 2024-04-09 Microsoft Technology Licensing, Llc Multi-content recommendation system combining user model, item model and real time signals
US11347752B2 (en) * 2018-07-23 2022-05-31 Apple Inc. Personalized user feed based on monitored activities
CN109299370B (zh) * 2018-10-09 2022-03-01 中国科学技术大学 多对级个性化推荐方法
CN109614555A (zh) * 2018-12-27 2019-04-12 腾讯音乐娱乐科技（深圳）有限公司 生成排行榜的方法和装置
CN110598016B (zh) * 2019-09-11 2021-08-17 腾讯科技（深圳）有限公司 一种多媒体信息推荐的方法、装置、设备和介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150319181A1 (en) * 2014-04-30 2015-11-05 Twitter, Inc. Application Graph Builder
CN107609146A (zh) * 2017-09-21 2018-01-19 北京达佳互联信息技术有限公司 信息展示方法、装置、终端及服务器
CN108334536A (zh) * 2017-11-30 2018-07-27 中国电子科技集团公司电子科学研究院 一种信息推荐方法、设备和存储介质
CN109902708A (zh) * 2018-12-29 2019-06-18 华为技术有限公司 一种推荐模型训练方法及相关装置
CN111339327A (zh) * 2020-02-20 2020-06-26 北京达佳互联信息技术有限公司 作品推荐方法及装置、服务器和存储介质

Also Published As

Publication number Publication date
JP2023512692A (ja) 2023-03-28
CN111339327A (zh) 2020-06-26
CN111339327B (zh) 2024-05-07
US20220398277A1 (en) 2022-12-15

Similar Documents

Publication Publication Date Title
WO2021164643A1 (zh) 作品推荐方法及服务器
CN108334887B (zh) 一种用户选取方法和装置
CN108306810B (zh) 一种群组组建方法、服务器及终端
CN113609392B (zh) 一种内容推荐方法、待推荐内容确定方法和相关装置
US11501327B2 (en) Method, apparatus, and system for delivering promotion information, and storage medium
WO2015090137A1 (en) A voice message search method, device, and system
CN110609957B (zh) 一种进行全局搜索的方法、终端和服务器
CN111143543A (zh) 一种对象推荐方法、装置、设备及介质
CN111125523B (zh) 搜索方法、装置、终端设备及存储介质
WO2016015471A1 (zh) 一种预测用户离网的方法及装置
CN104066009A (zh) 节目识别方法、装置、终端、服务器及系统
CN111259200A (zh) 视频类别划分方法、装置、电子设备及存储介质
CN111818350A (zh) 展示信息的方法、系统、装置、电子设备及存储介质
US20190223141A1 (en) Device positioning method and apparatus
CN111625690A (zh) 一种对象推荐方法、装置、设备及介质
CN113537685A (zh) 一种数据处理方法和装置
CN110796034B (zh) 一种目标对象识别方法、装置、设备及介质
CN114840565A (zh) 抽样查询方法、装置、电子设备及计算机可读存储介质
CN111797308A (zh) 一种资源推荐方法、装置、电子设备及介质
CN111191143B (zh) 应用推荐方法及装置
CN110287398B (zh) 一种信息更新的方法以及相关装置
CN109739409B (zh) 一种批处理的方法、装置以及终端设备
CN114430504B (zh) 一种媒体内容的推荐方法以及相关装置
CN112862289B (zh) 一种临床研究从业者的信息匹配方法和装置
CN114648336A (zh) 人脸支付方法、装置、电子设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21757012

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022547740

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21757012

Country of ref document: EP

Kind code of ref document: A1