WO2022247671A1 - 用户召回方法、装置、计算机设备和存储介质 - Google Patents

用户召回方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2022247671A1
WO2022247671A1 PCT/CN2022/092996 CN2022092996W WO2022247671A1 WO 2022247671 A1 WO2022247671 A1 WO 2022247671A1 CN 2022092996 W CN2022092996 W CN 2022092996W WO 2022247671 A1 WO2022247671 A1 WO 2022247671A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
recall
scheme
new
weight
Prior art date
Application number
PCT/CN2022/092996
Other languages
English (en)
French (fr)
Inventor
王宝川
曾涛
Original Assignee
百果园技术(新加坡)有限公司
王宝川
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 百果园技术(新加坡)有限公司, 王宝川 filed Critical 百果园技术(新加坡)有限公司
Publication of WO2022247671A1 publication Critical patent/WO2022247671A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/43Querying
    • G06F16/432Query formulation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/43Querying
    • G06F16/435Filtering based on additional data, e.g. user or group profiles
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/953Querying, e.g. by the use of web search engines
    • G06F16/9535Search customisation based on user profiles and personalisation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks

Definitions

  • the present application relates to the technical field of computer processing, for example, to a user recall method, device, computer equipment and storage medium.
  • multimedia data With the rapid development of science and technology, especially the widespread popularization of mobile communication networks and mobile terminals, a variety of multimedia data appear on the network, and many users can produce multimedia data anytime and anywhere. Different manufacturers provide platforms. On the one hand, collect some The multimedia data produced by users, on the other hand, screen high-quality multimedia data in different ways and push them to another part of users for display, reducing the difficulty for other part of users to obtain high-quality multimedia data.
  • the platform usually screens some high-quality users who produce multimedia data through a specific model, and pushes the user who produces multimedia data to the user who consumes multimedia data.
  • the application proposes a user recall method, device, computer equipment and storage medium to solve the problem of poor screening effect of users who produce multimedia data.
  • This application provides a user recall method, including:
  • the recalled user information of the second user is pushed to the client for display.
  • the application also provides a user recall device, including:
  • the user group determination module is configured to determine the user group where the first user for consuming multimedia data is located when receiving the request sent by the client, wherein the user group is associated with multiple recalls configured with weights Program;
  • the recall plan selection module is configured to select a recall plan for the first user according to a plurality of weights respectively corresponding to the plurality of recall plans;
  • a recall scheme execution module configured to execute the recall scheme for the first user, and respectively recall the second users used to produce multimedia data from a plurality of user pools;
  • the user information pushing module is configured to push the recalled user information of the second user to the client for display.
  • the present application also provides a kind of computer equipment, and described computer equipment comprises:
  • processors one or more processors
  • memory configured to store one or more programs
  • the one or more processors implement the above user recall method.
  • the present application also provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the above user recall method is implemented.
  • FIG. 1 is a flowchart of a user recall method provided in Embodiment 1 of the present application;
  • FIG. 2 is a schematic diagram of pushing user information provided in Embodiment 1 of the present application.
  • FIG. 3 is a flowchart of a user recall method provided in Embodiment 2 of the present application.
  • FIG. 4 is a schematic diagram of an iterative optimization recall scheme provided in Embodiment 2 of the present application.
  • FIG. 5 is a schematic structural diagram of a user recall device provided in Embodiment 3 of the present application.
  • FIG. 6 is a schematic structural diagram of a computer device provided in Embodiment 4 of the present application.
  • Figure 1 is a flow chart of a user recall method provided in Embodiment 1 of the present application.
  • This embodiment is applicable to the situation of dividing user groups to configure recall schemes and improving recall diversity.
  • This method can be executed by a user recall device , the user recall device can be realized by software and/or hardware, and can be configured in a computer device of a multimedia platform, such as a server, a workstation, a personal computer, etc., and the user recall method includes:
  • multimedia data On the multimedia platform, a large amount of multimedia data is usually stored to form a multimedia library.
  • the format of these multimedia data can include video data, audio data, image data, text data, etc., and its form is different in different business scenarios. For example, live shows, short videos, songs, comics, audiobooks, novels, news, articles, etc.
  • users registered on the multimedia platform can use electronic devices such as mobile terminals and personal computers to produce multimedia data and upload multimedia data to the multimedia platform.
  • the forms of multimedia data production are different.
  • users can use cameras to collect video data, and perform operations such as beautifying and connecting to the video data to generate live programs.
  • the short video business scenario users can use the camera to collect video data, perform operations such as beautifying, editing, and adding special effects on the video data to generate short videos.
  • users can use the keyboard to edit text and use images Editing tools Edit pictures, and typesetting text and pictures to generate news, etc.
  • the multimedia platform After verifying the legality of the multimedia data, the multimedia platform stores the multimedia data in the multimedia database and publishes the multimedia data. Other users registered on the multimedia platform can consume the multimedia data using electronic devices such as mobile terminals and personal computers, for example, Watch live shows, short videos, listen to songs, audiobooks, read news, articles, and more.
  • electronic devices such as mobile terminals and personal computers, for example, Watch live shows, short videos, listen to songs, audiobooks, read news, articles, and more.
  • the user who consumes multimedia data can be called the first user, and the user who produces multimedia data can be called the second user.
  • the registered users can both produce multimedia data and consume multimedia data. , that is, the user is both the first user and the second user, which is not limited in this embodiment.
  • the first user and the second user can be natural persons, groups, studios, enterprises and institutions, etc., which are not limited in this embodiment.
  • the meanings of the first user and the second user are also different.
  • the second user is the broadcaster and the first user is the audience.
  • the second user The second user is an editor, and the first user is a reader.
  • the second user is a singer, the first user is a fan, and so on.
  • the first user can use a user account to log in to the multimedia platform in the client.
  • the client can be a browser, an independent multimedia application (such as a live application, a short video application, a player application, a news application, etc.), or a Small programs loaded in applications (such as instant messaging tools, payment applications, etc.), etc., the client can send a request to the multimedia platform, and the multimedia platform triggers the operation of screening the second user for the first user when receiving the request from the client .
  • the client's request can be triggered by the first user.
  • the first user enters a keyword on the client and requests the multimedia platform to search for the second user related to the keyword, and the first user pulls down the list request of the existing second user.
  • the multimedia platform refreshes the second user, and so on.
  • the request of the client may not be triggered by the first user.
  • the multimedia platform is requested to push the relevant second user, etc., which is not limited in this embodiment.
  • the plan for recalling the second user i.e., the recall plan
  • the recall plan relies on more data for optimization, so as to obtain relatively stable indicators. Therefore, multiple user groups can be set in advance.
  • a user has at least one of the same attributes, set a recall plan for user groups, and use the recall plan of user groups instead of individual users as the optimization target, thereby increasing the number of indicators for later feedback and improving the stability of indicators, thereby improving The quality of the recall protocol.
  • the behavior of the first user is counted every certain time interval (such as 3 days), the first user is grouped, the first user is divided into the user group, and the first user (in the form of user Identifier (Identifier, ID) and other data identification) and the mapping relationship between the user group, when the request sent by the client is received online, the user group of the first user can be read in the mapping relationship.
  • the user group may take pushing the information of the second user to the first user as the ultimate goal to be optimized, and different user groups may be divided according to different goals.
  • S101 includes:
  • S1011 Collect statistics on the behavior data generated by the consumption of multimedia data by the first user within each first time period.
  • a longer second time period (such as 60 days) can be divided into multiple shorter first time periods (such as 1 day), and in each first time period, the first time period can be counted
  • a user's behavioral data generated when consuming multimedia data for example, the number of multimedia data browsed, the duration of multimedia data browsed, the fees used to pay for value-added services (such as subscriptions, virtual gifts, etc.), and so on.
  • the first user By analyzing the behavior data generated when the first user triggers the multimedia data within the first time period, it can be known that the first user is interested in the multimedia data in some dimensions (such as cost consumption, paying attention to the second user, etc.) during the first time period.
  • Presented behavioral expressions are tagged as attributes to the first time period.
  • Behavior data can be quantified, and one or more behavior conditions can be set for different attributes. If the behavior data of a first time period satisfies one of the behavior conditions, the attribute corresponding to the behavior condition can be marked to the first time period.
  • the attributes include effective time and effective relationship chain time, where the effective time indicates the time for effectively consuming multimedia data, and the effective relationship chain time indicates the time for effectively consuming multimedia data with a concern relationship.
  • the first time threshold (behavior condition) If the total time for the first user to consume multimedia data within the first time period is greater than the first time threshold (behavior condition), mark the attribute as valid time for the first time period.
  • the first time threshold (behavior condition) If the total time spent by the first user consuming the target data within the first time period is greater than or equal to the first time threshold (behavior condition), mark the first time period as valid relationship chain time, where the target data is the first user Multimedia data produced by the concerned second user.
  • the first frequency of occurrence of the effective time within the second time period may be counted, and the second frequency of occurrence of the effective relationship chain time within the second time period may be counted, Calculates the ratio between the second frequency and the first frequency.
  • the first frequency is less than the first frequency threshold (for example, 5 days)
  • the first frequency threshold for example, 5 days
  • the first frequency threshold such as 5 days
  • the second frequency threshold such as 15 days
  • the ratio is smaller than the first ratio threshold (such as 50%)
  • the first frequency is greater than the first frequency threshold (such as 5 days) and less than the second frequency threshold (such as 15 days), and the ratio is greater than the first ratio threshold (such as 50%), then determine the user group to which the first user belongs Groups are low-relationship links and high-efficiency groups.
  • the first frequency is greater than the second frequency threshold (such as 15 days), and the ratio is smaller than the first ratio threshold (such as 50%), then it is determined that the user group to which the first user belongs is a high-validity-low-relationship chain group.
  • the first frequency is greater than the second frequency threshold (such as 15 days), and the ratio is greater than the first ratio threshold (such as 50%), then it is determined that the user group to which the first user belongs is a group with high validity and high relationship chain.
  • the user groups are divided according to whether they effectively consume multimedia data and whether they effectively consume multimedia data produced by a second user who has been followed, so as to optimize the first user who gets high attention and/or the first user who consumes most.
  • user groups are only examples. When implementing the embodiments of the present application, other user groups can be set according to the actual situation. Groups, etc., are not limited in this embodiment of the present application.
  • a user group may be associated with multiple (at least two) recall schemes configured with weights, in which recall schemes are used to record the elements of recalling the second user from the user pool, and for each recall scheme, Configure the weight.
  • the weight takes a value between (0, 1).
  • the sum of the weights of all recall schemes in the same user group is 1.
  • the weight corresponding to the recall scheme represents the recall scheme at the time of allocation.
  • the index of the first user feedback is the better recall plan, its weight is higher, and it can get a higher chance of exposure.
  • a recall plan with a worse indicator has a lower weight, can get a lower chance of exposure, and is distributed to the first user in the user group for fewer times.
  • the weight of each recall scheme can be mapped to a probability, where the probability is positively correlated with the weight, that is, the larger the weight, the higher the probability, and the smaller the weight, the lower the probability.
  • the weight may be converted into a percentage as a probability.
  • one recall plan among the plurality of recall plans may be discretely selected for the first user according to multiple probabilities respectively corresponding to the multiple recall plans.
  • the probability of each recall scheme can be converted into adjacent sub-ranges, and a floating-point number is randomly generated in the total range composed of the sub-ranges, thereby selecting the recall scheme corresponding to the sub-range where the floating-point number is located , assigned to the first user.
  • the subrange of is [0, 0.5), The subrange of [0.5, 0.75), The sub-range of is [0.75, 1], a floating-point number is randomly generated between the total range [0, 1], if the floating-point number is within [0, 0.5), it is selected Selected if the float is in [0.5, 0.75) Selected if the float is within [0.75, 1]
  • the user pool includes at least one second user as follows:
  • the second user who is easy to be followed can be recalled, which can improve the relationship chain behavior of the first user, for example, the recalled second user increases the frequency of effective relationship chain time.
  • recalling the second users who are highly consumed can increase the consumption time and increase the number of first users who trigger effective consumption behaviors.
  • a certain proportion such as more than 15 days
  • the preset time threshold such as 10 minutes
  • a neural network, machine learning, etc. may be applied to train a personalized recommendation model, and these personalized models may be recalled for a second user that is personally matched by the first user.
  • the online indicators of the second user can be mapped to popularity through weighted summation, and the second users can be sorted according to popularity , take out a certain proportion (such as the first 5%) of the second users and put them into the user pool, and when the second users in the user pool are offline, the number of second users is guaranteed to be recalled.
  • the above user pool is just an example.
  • other user pools may be set according to the actual situation, which is not limited in the embodiment of the present application.
  • the recall scheme may include multiple elements, the position of the element represents the user pool that recalls the second user (that is, there is a mapping relationship between the position of the element and the user pool (identified by data such as ID)), and the element The value of represents the number of second users recalled from the user pool.
  • the element configured in each user pool can be read for the first user in the recall scheme, and the element has a position and a value, so that the first user can recall the number of items from the user pool corresponding to the position The second user for the value.
  • the recall scenario is in, There are 4 elements in total.
  • the first element corresponds to user pool A
  • the second element corresponds to user pool B
  • the third element corresponds to user pool C
  • the fourth element corresponds to user pool D.
  • the process of pushing the user information of the second user usually includes the following operations:
  • the second user is recalled from the user pool, and the set of optional video data is narrowed down.
  • the number of recalled second users is relatively large, usually on the order of tens of thousands or thousands, and the algorithm used for fine sorting may be more complicated.
  • a rough sorting link can be added between recall and fine sorting.
  • Load the number of recalled second users into a simple ranking model through a small number of second users and features of multimedia data, such as Logistic Regression (LR) model, Gradient Boost Decision Tree (GBDT) Model, etc. roughly sort the recalled second users, select the second users with higher rankings, and reduce the number of fine-sorted second users under the premise of ensuring a certain degree of accuracy.
  • the second users can be The number dropped to the order of thousands or hundreds.
  • rough sorting is often optional, that is, rough sorting can be applied, and it can also be directly switched from recall to fine sorting, which is not limited in this embodiment.
  • the number of second users roughly sorted is loaded into a more complex ranking model, for example, convolutional neural network (Convolutional Neural Networks, CNN), recurrent neural network (Recurrent Neural Network, RNN), etc., accurately sort the rough sorted video data, select the second users with higher ranking, try to improve the accuracy of sorting, and reduce the number of second users sent to the client. Reduce the number of second users to the order of one hundred or ten.
  • convolutional neural network Convolutional Neural Networks, CNN
  • Recurrent Neural Network Recurrent Neural Network
  • the second users extracted after fine sorting can be called a user sequence, and after breaking up (also called rearranging) the second users in the user sequence, the number of second users is maintained (such as hundred, ten), send the user information (such as name, avatar, etc.) of the second user to the client, and the client displays the user information of multiple second users in order for the first user to browse.
  • the user information of the second user can be triggered by mouse click, touch click, etc., and the client requests the second user's page from the multimedia platform, for example, the second user A page hosting a live program, a page aggregating multimedia data released by a second user, etc., is loaded for the second user to browse.
  • the user group of the first user for consuming multimedia data is determined, and the user group is associated with multiple recall schemes configured with weights, and the user group
  • the user group is associated with multiple recall schemes configured with weights, and the user group
  • it can increase the amount of data that users feedback on the same recall plan, and can reduce the situation that a single user’s feedback on the recall plan is too small to cause disbelief, even if the data fed back by a single user fluctuates or Noise is also acceptable
  • select a recall plan for the first user according to multiple weights corresponding to multiple recall plans execute the recall plan for the first user, and recall the second user for producing multimedia data from multiple user pools , the user information of the recalled second user is pushed to the client for display.
  • recalling the second user from different user pools makes different user groups Groups can have appropriate quotas, which increases the diversity of recalls, improves the recall of second users who are interested in the first user, and facilitates the iteration of the recall plan in the later stage to optimize the recall plan.
  • the first user can directly trigger the user information of the interested second user without searching for the interested second user through keywords, which greatly improves the convenience of obtaining interested information and reduces the need to recall the second user and push the second user on the multimedia platform.
  • the user information of the second user, and the device of the first user display the resources of the user information of the second user.
  • Fig. 3 is a flow chart of a user recall method provided in Embodiment 2 of the present application. This embodiment is based on the foregoing embodiments and adds an operation of updating the recall plan. The method includes:
  • user groups are associated with multiple recall schemes with weighted configurations.
  • the user groups include multiple first users. Considering the real-time update of recall schemes, the historical data of the first users are not used. Therefore, the user group When the first user in the group changes slightly, the recall scheme of the user group is generally not reinitialized.
  • a recall scheme for a user group can be initialized.
  • the user group determines the upper limit N of the number of recall schemes in the user group G, and the total number M of recalled second users in each recall scheme.
  • Configure N recall schemes for user group G The set formed by the mapping to the weight W, namely ensure (i.e., the recall protocol).
  • the sum of the numerical values of multiple elements in is the total number M).
  • the numerical values of multiple elements in the recall scheme are the same, and the weights of all recall schemes are the same and the sum is 1.
  • technicians can update the user pool by optimizing the results of the recall plan, such as adding a user pool, deleting a user pool, etc., to make the entire recall process more efficient and assist in the iteration of the recall plan.
  • each recall scheme using the user pool for recall is updated accordingly.
  • the recall scheme includes a plurality of elements, the positions of the elements represent the user pools for recalling the second users, and the values of the elements represent the number of recalled second users.
  • the length of the recall plan is increased by one, and an element indicating the user pool is added in the recall plan, and the mapping relationship between the position of the newly added element and the newly added user pool is recorded, and the newly added
  • the value of the element is set to the average value of the values of other elements. If the total number of recalled second users in multiple recall schemes is limited, the values of the current multiple elements can be scaled proportionally and rounded up (including rounding up) Integer, rounded down), so that the sum of the values of multiple elements in the recall scheme is the total.
  • the length of the recall plan is reduced by one, and the element indicating the user pool is deleted in the recall plan, that is, according to the mapping relationship, find the mapping position of the user pool to be deleted, and delete the element at this position.
  • the total number of second users recalled in a recall plan the values of the current multiple elements can be scaled proportionally and rounded (including rounding up and rounding down), so that the values of multiple elements in the recall plan The sum of the values is the total.
  • the recall plan is executed, and the second user is recalled from the corresponding user pool according to the configuration of the recall plan, and the user information of the second user is pushed to the client for browsing by the first user.
  • the push of the user information of the second user is real-time. Therefore, the index for evaluating the second user is real-time. Since a user group is used as an optimization target instead of a single user, it is acceptable even if a single user evaluates the second user's index with fluctuations or noises.
  • the indicator is the first The situation when the user clicks on the user information of the second user (such as the amount of clicks, click rate), the duration of consuming the multimedia data produced by the second user, whether the first user pays attention to the second user, etc.
  • These indicators can be used individually or in combination to generate a final indicator, for example, weighted average of multiple indicators, multiplication after normalization of multiple indicators, etc., which is not limited in this embodiment.
  • the recall plan can be gradually optimized, and the user group can be carried out in the desired direction.
  • the indicators generated by the first user under each recall scheme can reflect the quality of the recall scheme to a certain extent, and can be used as a basis to update the weight of the recall scheme.
  • the first user feedbacks more positive indicators, indicating that the quality of the recall plan is better. At this time, the weight of the recall plan can be increased. The first user feedbacks more negative indicators, indicating that the quality of the recall plan is worse. At this time, the weight of the recall scheme can be decreased.
  • the number of first users delivered or the number of indicators obtained will be different. If a recall plan sends a small number of first users , may lead to untrustworthy indicators due to fluctuations.
  • the relative advantages and disadvantages of the recall scheme can be measured without estimating the expected index size of the recall scheme. Therefore, all the recall schemes can be sorted according to the size of the index, and multiple Indexes are divided into the first candidate index and the second candidate index, wherein the first candidate index is higher than the second candidate index, taking 50% as an example, the first candidate index is the 50% index with the highest value, and the second candidate index 50% of indicators with the lowest value.
  • the first candidate indicator Take the first candidate indicator as a positive example and the second candidate indicator as a negative example, count the number of positive examples and the number of negative examples, and use Wilson interval method, normal interval (Normal approximation interval) and other methods to calculate the confidence interval for the recall plan , so that the weight is corrected based on the confidence interval, and the revised weight is used as the new weight of the recall scheme.
  • Wilson interval method Normal interval (Normal approximation interval) and other methods to calculate the confidence interval for the recall plan , so that the weight is corrected based on the confidence interval, and the revised weight is used as the new weight of the recall scheme.
  • the weights may be modified in a smooth manner.
  • a first product between the lower value of the confidence interval and the first coefficient can be calculated.
  • the essence of the confidence interval is to correct the credibility to make up for the influence of the small sample size (ie, the index of the first user feedback). If the sample size is large, it means that it is more credible and less corrections are made. At this time, the confidence interval will be relatively narrow and the lower limit value will be relatively large; At this time, the confidence interval will be relatively wide, and the lower limit value will be relatively small.
  • the lower limit of the confidence interval is calculated as follows:
  • u is a positive example
  • v is a negative example
  • z ⁇ is a constant related to the confidence interval, which can be 2
  • Score is the lower limit of the confidence interval.
  • a second product may be calculated between the weight and a second coefficient, wherein the second coefficient is greater than the first coefficient to account for occasional fluctuations.
  • W is the updated weight
  • W new is the lower limit of the confidence interval
  • W old is the original weight
  • is the second coefficient, such as 0.25
  • is the first coefficient, such as 0.75.
  • the genetic algorithm for the recall scheme after updating the weights, the genetic algorithm (Genetic Algorithm, GA) can be used to perform genetic operations on the recall scheme and its weight, and inherit the genes of the relatively high-quality recall scheme, thereby generating a new Recall scheme, configure new weights for the new recall scheme.
  • GA Genetic Algorithm
  • Genetic algorithm takes all individuals (recall scheme) in a population (recall scheme of the same user group) as objects, and uses randomization technology to guide an efficient search of an encoded parameter space. Among them, selection, crossover and mutation constitute the genetic operation of genetic algorithm.
  • the offspring population is more adaptable to the environment than the previous generation.
  • the optimal individual in the last generation population can be used as an approximate optimal solution to the problem after decoding.
  • S307 includes:
  • the recall scheme can be selected, that is, all the recall schemes are sorted according to the size of the weight, and multiple recall schemes are divided into the first target scheme according to a certain ratio, and
  • the second objective scheme makes the weight of the first objective scheme higher than the weight of the second objective scheme. Taking 50% as an example, the first objective scheme is the 50% recall scheme with the highest weight, and the second objective scheme is the 50% with the lowest value. % recall scenarios.
  • a cross operation can be performed on some of the first target solutions.
  • any first target solution can be selected to exchange some elements, and the first target solution after the interaction element is
  • the first target solutions participating in the crossover operation can be sorted according to their weights, the sorted first target solutions are traversed, and for the current first target solution, other first target solutions with higher weight than the current first target solution are searched.
  • a usergroup with (with a weight of 0.4), (with a weight of 0.3), (with a weight of 0.2), (with a weight of 0.1) there are four first objective schemes, then it can be and Cross-operate between and and Cross-operate between and and and cross operation between them.
  • the sum of the values of multiple elements in the new recall scheme generated based on the crossover operation is not necessarily equal to the total number.
  • the values of multiple elements can be proportional Perform scaling and rounding operations (including rounding up and rounding down), so that the sum of the values of multiple elements in the recall scheme is the total.
  • the exposure chance of the new recall scheme can be reduced, and the smaller weight of the first target scheme that will participate in the crossover operation (that is, the weight of the current first target scheme ) is assigned to the weight of the new recall plan. If the quality of the new recall plan is better, its weight can be increased later when updating the weight. If the quality of the new recall plan is poor, less exposure opportunities can reduce the push The effect of the second user's user information is affected.
  • a mutation operation can be performed on part of the first target plan.
  • any first target plan can be selected to change some elements, and the first target plan after the element change is the user group A new recall scheme is formed, and a new weight is configured for the new recall scheme with reference to the weight of the first target scheme participating in the change operation.
  • the first target plan that participates in the crossover operation does not participate in the mutation operation. If the number N of recall plans for the same user group is limited, the mutation operation can be continuously performed on the first target plan until all recall The number of schemes reaches the upper limit N, if all the first target schemes execute the mutation operation and still do not reach the upper limit N, then the recall scheme generated in the previous mutation operation can be used as the new first target operation to perform the mutation operation again until all The number of recall scenarios reaches an upper limit N.
  • the first target solutions participating in the mutation operation can be sorted according to their weights, the sorted first target solutions can be traversed, and the value interval can be determined based on the values of the elements of the first target solution for the current first target solution , the lower limit of the numerical interval may be greater than, less than or equal to the smallest numerical value among the elements of the first target scheme, and the upper limit of the numerical interval may be greater than, less than or equal to the largest numerical value among the elements of the first target scheme, for example , subtract one from the smallest value among the elements of the first target solution as the lower limit value of the value range, and add one to the maximum value among the elements of the first target solution as the upper limit value of the value range.
  • the sum of the values of multiple elements in the new recall scheme generated based on the mutation operation is not necessarily equal to the total number.
  • the values of multiple elements can be proportional Perform scaling and rounding operations (including rounding up and rounding down), so that the sum of the values of multiple elements in the recall scheme is the total.
  • the exposure chance of the new recall scheme can be reduced, and the weight of the first target scheme is given a specified proportion (the proportion is less than 1, such as 80%), as the new The weight of the recall plan. If the quality of the new recall plan is good, you can increase its weight when updating the weight later. If the quality of the new recall plan is poor, less exposure opportunities can reduce the user who pushes the second user The impact of the effect of information.
  • the weights of all the recall schemes of the same user group are limited to 1, at this time, the weights of all the recall schemes of the user group are normalized according to the proportion, so that the weights of all the recall schemes of the user group The sum is 1.
  • the above-mentioned genetic operation is just an example of the operation of optimizing the recall scheme.
  • other operations of optimizing the recall scheme can be set according to the actual situation, for example, particle swarm algorithm, ant algorithm, gradient descent method, etc., this The embodiment of the application does not limit this.
  • the operation of the above-mentioned real-time optimized recall scheme may be turned off.
  • the recall scheme with the highest weight can be selected as the final recall scheme, and the above-mentioned real-time optimized recall scheme can be re-enabled.
  • the initialization can be skipped, and the number of recall schemes of a user group reaches the upper limit N through continuous mutation operations, and then the recall scheme is optimized according to the process.
  • statistics are generated by the first user under each recall scheme for evaluating the second user's index, and the weight of each recall scheme is updated according to the index, and the updated weight is used on the basis of the recall scheme as
  • the user group generates a new recall plan and configures a new weight for the new recall plan. Since the index for evaluating the second user can be obtained in real time, and the operation of updating the weight and iterative optimization is relatively simple, the optimization of the recall plan can be Guaranteed real-time performance, greatly improving the fit between the recall plan and the first user.
  • this embodiment updates the weight of the recall plan according to the first user’s feedback on the recall plan, thereby Acting on the optimization of the recall plan, it can improve the effect of the recall and avoid the unsatisfactory recall effect caused by artificial design.
  • optimizing the recall plan a better user pool has more room to play, and a poor user pool will Gradually eliminated, so that the quality of the recall program can be improved.
  • Fig. 5 is a structural block diagram of a user recall device provided in Embodiment 3 of the present application, which may include the following modules:
  • the user group determination module 501 is configured to, when receiving the request sent by the client, determine the user group in which the first user for consuming multimedia data belongs, and the user group is associated with multiple recall schemes configured with weights
  • the recall plan selection module 502 is configured to select a recall plan for the first user according to a plurality of weights respectively corresponding to a plurality of recall plans;
  • the recall plan execution module 503 is configured to execute the recall plan,
  • the second users used for producing multimedia data are respectively recalled from multiple user pools;
  • the user information pushing module 504 is configured to push the user information of the recalled second users to the client for display.
  • the user recall device provided in the embodiment of the present application can execute the user recall method provided in any embodiment of the present application, and has corresponding functional modules and effects for executing the method.
  • FIG. 6 is a schematic structural diagram of a computer device provided in Embodiment 4 of the present application.
  • FIG. 6 shows a block diagram of an exemplary computer device 12 suitable for implementing embodiments of the present application.
  • Components of computer device 12 may include, but are not limited to, one or more processors or processing units 16, memory 28, bus 18 connecting various system components including memory 28 and processing unit 16.
  • Computer device 12 includes a variety of computer system readable media. These media can be any available media that can be accessed by computer device 12 and include both volatile and nonvolatile media, removable and non-removable media.
  • Memory 28 may include computer system readable media in the form of volatile memory, such as Random Access Memory (RAM) 30 and/or cache memory 32 .
  • Storage system 34 may be used to read and write to non-removable, non-volatile magnetic media (not shown in FIG. 6, commonly referred to as a "hard drive").
  • the memory 28 may include at least one program product having a set (eg, at least one) of program modules configured to perform the functions of the embodiments of the present application.
  • Program/utility 40 having a set (at least one) of program modules 42, may be stored in memory 28, for example.
  • the program modules 42 generally perform the functions and/or methods of the embodiments described herein.
  • Computer device 12 may also communicate with one or more external devices 14 (eg, keyboard, pointing device, display 24, etc.). Such communication may occur through input/output (I/O) interface 22 .
  • the computer device 12 can also be connected to one or more networks (such as a local area network (Local Area Network, LAN), a wide area network (Wide Area Network, WAN) and/or a public network through the network adapter 20.
  • networks such as a local area network (Local Area Network, LAN), a wide area network (Wide Area Network, WAN) and/or a public network through the network adapter 20.
  • the processing unit 16 executes various functional applications and data processing by running the programs stored in the memory 28, for example, implementing the user recall method provided in the embodiment of the present application.
  • Embodiment 5 of the present application also provides a computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, multiple processes of the above-mentioned user recall method can be realized, and the same technical Effect, in order to avoid repetition, will not repeat them here.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Information Transfer Between Computers (AREA)
  • Two-Way Televisions, Distribution Of Moving Picture Or The Like (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

提供了一种用户召回方法、装置、计算机设备和存储介质。该用户召回方法包括:当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组(S101),用户群组关联多个配置有权重的召回方案;根据多个召回方案分别对应的多个权重为第一用户选择召回方案(S102);为第一用户执行召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户(S103);将召回的第二用户的用户信息推送至客户端进行显示(S104)。

Description

用户召回方法、装置、计算机设备和存储介质
本申请要求在2021年05月24日提交中国专利局、申请号为202110563301.5的中国专利申请的优先权,该申请的全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机处理的技术领域,例如涉及一种用户召回方法、装置、计算机设备和存储介质。
背景技术
随着科技的快速发展,尤其是移动通信网络与移动终端的广泛普及,网络上出现了多种多样的多媒体数据,许多用户可以随时随地生产多媒体数据,不同的厂商提供平台,一方面,收集一部分用户生产的多媒体数据,另一方面,通过不同的方式筛选优质的多媒体数据并推送至另一部分用户进行显示,降低另一部分用户获取优质多媒体数据的难度。
由于一个优质用户生产的多媒体数据具有质量的保证,因此,平台通常是通过特定的模型筛选一些优质的、生产多媒体数据的用户,将该生产多媒体数据的用户推送至消费多媒体数据的用户。
通常是通过简单的条件召回生产多媒体数据的用户,召回的效果较差,导致筛选生产多媒体数据的用户的效果较差,此时,消费多媒体数据的用户往往忽略推送过来的用户,通过关键词搜索感兴趣的用户,这些操作较为繁琐、浪费较多的资源。
发明内容
本申请提出了一种用户召回方法、装置、计算机设备和存储介质,以解决筛选生产多媒体数据的用户的效果较差的问题。
本申请提供了一种用户召回方法,包括:
当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组,其中,所述用户群组关联多个配置有权重的召回方案;
根据多个召回方案分别对应的多个权重为所述第一用户选择召回方案;
为所述第一用户执行所述召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户;
将召回的第二用户的用户信息推送至所述客户端进行显示。
本申请还提供了一种用户召回装置,包括:
用户群组确定模块,设置为当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组,其中,所述用户群组关联多个配置有权重的召回方案;
召回方案选择模块,设置为根据多个召回方案分别对应的多个权重为所述第一用户选择召回方案;
召回方案执行模块,设置为为所述第一用户执行所述召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户;
用户信息推送模块,设置为将召回的第二用户的用户信息推送至所述客户端进行显示。
本申请还提供了一种计算机设备,所述计算机设备包括:
一个或多个处理器;
存储器,设置为存储一个或多个程序;
当所述一个或多个程序被所述一个或多个处理器执行,使得所述一个或多个处理器实现上述的用户召回方法。
本申请还提供了一种计算机可读存储介质,所述计算机可读存储介质上存储计算机程序,所述计算机程序被处理器执行时实现上述的用户召回方法。
附图说明
图1为本申请实施例一提供的一种用户召回方法的流程图;
图2为本申请实施例一提供的一种推送用户信息的示意图;
图3是本申请实施例二提供的一种用户召回方法的流程图;
图4是本申请实施例二提供的一种迭代优化召回方案的示意图;
图5为本申请实施例三提供的一种用户召回装置的结构示意图;
图6为本申请实施例四提供的一种计算机设备的结构示意图。
具体实施方式
下面结合附图和实施例对本申请进行说明。此处所描述的具体实施例仅仅用于解释本申请。为了便于描述,附图中仅示出了与本申请相关的部分。
实施例一
图1为本申请实施例一提供的一种用户召回方法的流程图,本实施例可适用于划分用户群组配置召回的方案、提高召回多样性的情况,该方法可以由用户召回装置来执行,该用户召回装置可以由软件和/或硬件实现,可配置在多媒体平台的计算机设备中,例如,服务器、工作站、个人电脑,等等,该用户召回方法包括:
S101、当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组。
在多媒体平台上,通常存储大量的多媒体数据,形成多媒体库,这些多媒体数据的格式可以包括视频数据、音频数据、图像数据、文本数据,等等,其形式在不同的业务场景中有所不同,例如,直播节目、短视频、歌曲、漫画、有声小说、小说、新闻、文章,等等。
除了从版权方购买多媒体数据之外,在多媒体平台注册的用户可以使用移动终端、个人电脑等电子设备生产多媒体数据,将多媒体数据上传至多媒体平台。
对于不同的业务场景,生产多媒体数据的形式有所不同,例如,在直播的业务场景中,用户可以使用摄像机采集视频数据,对视频数据进行美颜、连麦等操作,从而生成直播节目,在短视频的业务场景中,用户可以使用摄像机采集视频数据,对视频数据进行美颜、剪辑、添加特效等操作,从而生成短视频,在新闻的业务场景中,用户可以使用键盘编辑文字、使用图像编辑工具编辑图片,并对文字、图片进行排版,从而生成新闻,等等。
多媒体平台在审核该多媒体数据合法之后,将该多媒体数据存储至多媒体库,并发布该多媒体数据,在多媒体平台注册的其他用户,可使用移动终端、个人电脑等电子设备消费该多媒体数据,例如,观看直播节目、短视频,收听歌曲、有声小说,阅读新闻、文章,等等。
为便于区分,消费多媒体数据的用户可称之为第一用户,生产多媒体数据的用户可称之为第二用户,对于一些多媒体平台,其注册的用户既可以生产多媒体数据、也可以消费多媒体数据,即该用户既是第一用户,也是第二用户,本实施例对此不加以限制。
一般情况下,该第一用户、第二用户均可以为自然人,也可以为团体、工作室、企事业单位,等等,本实施例对此不加以限制。
在不同的业务场景中,第一用户、第二用户的含义也有所不同,例如,在直播的业务场景中,第二用户为主播、第一用户为观众,在自媒体的业务场景中,第二用户为编辑、第一用户为读者,在音乐的业务场景中,第二用户为歌 手、第一用户为歌迷,等等。
第一用户可使用用户账号等方式在客户端中登录多媒体平台,该客户端可以为浏览器、独立的多媒体应用(如直播应用、短视频应用、播放器应用、新闻应用等)、在第三方应用(如即时通讯工具、支付应用等)中加载的小程序,等等,客户端可向多媒体平台发送请求,多媒体平台在接收到客户端的请求时,触发为第一用户筛选第二用户的操作。
客户端的请求可以由第一用户主动触发,例如,第一用户在客户端输入关键词并请求多媒体平台搜索与该关键词相关的第二用户,第一用户下拉已有的第二用户的列表请求多媒体平台刷新第二用户,等等,客户端的请求也可以并非第一用户主动触发,例如,客户端在显示主页时请求多媒体平台推送第一用户可能感兴趣的第二用户,客户端在当前多媒体数据结束播放前请求多媒体平台推送相关的第二用户,等等,本实施例对此不加以限制。
一般情况下,召回第二用户的方案(即召回方案)依赖较多的数据进行优化,从而得到相对稳定的指标,因此,可以预先设置多个用户群组,在同一个用户群组中的第一用户至少具有一个相同的属性,针对用户群组设置召回方案,使用用户群组而并非单个用户的召回方案作为优化的目标,从而提高后期反馈的指标的数量、提高指标的稳定性,从而提高召回方案的质量。
在离线时,每间隔一定的时间(如3天)统计一次第一用户的行为,对第一用户进行分组处理,将第一用户划分至该用户群组中,并记录第一用户(以用户标识(Identifier,ID)等数据标识)与用户群组之间的映射关系,当在线接收到客户端发送的请求时,可以在该映射关系中读取第一用户所处的用户群组。
用户群组可以以推送第二用户的信息给第一用户作为最终要优化的目标,针对不同的目标可以划分不同的用户群组。
在本申请的一个实施例中,S101包括:
S1011、统计第一用户在每个第一时间段内消费多媒体数据所产生的行为数据。
在本实施例中,可以将一个较长的第二时间段(如60天)划分为多个较短的第一时间段(如1天),在每个第一时间段中,可以统计第一用户在消费多媒体数据时产生的行为数据,例如,浏览多媒体数据的数量、浏览多媒体数据的时长、用于支付增值业务(如订阅、赠送虚拟礼物等)的费用,等等。
S1012、基于行为数据识别每个第一时间段的属性。
通过分析第一时间段内、第一用户触发多媒体数据时产生的行为数据,可以获知第一用户在该第一时间段内对多媒体数据在一些维度(如费用消耗、关 注第二用户等)下呈现的行为表达,将这些行为表达作为属性标记至该第一时间段。
行为数据可以进行数值化,针对不同属性设置一个或多个行为条件,若一个第一时间段的行为数据满足其中一个行为条件,则可以将该行为条件对应的属性标记至该第一时间段。
在一个示例中,属性包括有效时间、有效关系链时间,有效时间表示有效消费多媒体数据的时间,有效关系链时间表示有效消费具有关注关系的多媒体数据的时间。
若第一用户在第一时间段内消费多媒体数据的总时间大于第一时间阈值(行为条件),则对第一时间段标记属性为有效时间。
若第一用户在第一时间段内消费目标数据的总时间大于或等于第一时间阈值(行为条件),则对第一时间段标记属性为有效关系链时间,其中,目标数据为第一用户关注的第二用户生产的多媒体数据。
上述属性及其识别方式只是作为示例,在实施本申请实施例时,可以根据实际情况设置其它属性及其识别方式,本申请实施例对此不加以限制。
S1013、当多个第一时间段的属性在第二时间段内符合预设的群组条件时,则将第一用户划分至群组条件对应的用户群组。
针对同一个用户,可以对多个第一时间段在第二时间段内的多个属性进行统计,以数值化进行表示,针对不同用户群组设置一个或多个群组条件,若第二时间段内统计的属性满足其中一个群组条件,则标记第一用户属于该群组条件对应的用户群组。
在一个示例中,如果属性包括有效时间、有效关系链时间,则可统计有效时间在第二时间段内出现的第一频次、统计有效关系链时间在第二时间段内出现的第二频次,计算第二频次与第一频次之间的比值。
若第一频次小于第一频次阈值(如5天),则确定第一用户归属的用户群组为新用户群组或低活跃群组。
若第一频次大于第一频次阈值(如5天)、且小于第二频次阈值(如15天),并且,比值小于第一比例阈值(如50%),则确定第一用户归属的用户群组为低关系链低有效群组。
若第一频次大于第一频次阈值(如5天)、且小于第二频次阈值(如15天),并且,比值大于第一比例阈值(如50%),则确定第一用户归属的用户群组为低关系链高有效群组。
若第一频次大于第二频次阈值(如15天),并且,比值小于第一比例阈值(如50%),则确定第一用户归属的用户群组为高有效低关系链群组。
若第一频次大于第二频次阈值(如15天),并且,比值大于第一比例阈值(如50%),则确定第一用户归属的用户群组为高有效高关系链群组。
在本示例中,通过是否有效消费多媒体数据、是否有效消费已关注的第二用户生产的多媒体数据划分用户群组,可以优化得到高关注的第一用户、和/或高消费的第一用户。
上述用户群组只是作为示例,在实施本申请实施例时,可以根据实际情况设置其它用户群组,例如,按照注册的时间划分用户群组(区分新用户、老用户)、按照地理区域划分用户群组,等等,本申请实施例对此不加以限制。
S102、根据多个召回方案分别对应的多个权重为第一用户选择召回方案。
在本实施例中,一个用户群组可关联多个(至少两个)配置有权重的召回方案,该召回方案中用于记录从用户池召回第二用户的元素,对于每一个召回方案,可配置权重,一般情况下,权重在(0,1)之间取值,在同一个用户群组内所有召回方案的权重之和为1,召回方案对应的权重代表了这个召回方案在分配时的优先度,第一用户反馈的指标更优的召回方案,其权重更高,可以得到更高的曝光机会,下发给该用户群组中第一用户使用的次数越多,第一用户反馈的指标更差的召回方案,其权重更低,可以得到更低的曝光机会,下发给该用户群组中第一用户使用的次数越少。
在实现中,可将每个召回方案的权重映射为概率,其中,概率与权重正相关,即权重越大,概率越高,权重越小,概率越低。示例性地,如果在同一个用户群组内所有召回方案的权重之和为1,则可以将该权重转换为百分比的形式,作为概率。
若映射概率完成,则可以根据多个召回方案分别对应的多个概率离散地为第一用户选择多个召回方案中的一个召回方案。
在一种选择方式中,可以将每个召回方案的概率转换为相邻的子范围,在子范围组成的总范围内随机生成一浮点数,从而选定该浮点数所在子范围对应的召回方案、分配给第一用户。
示例性地,假设针对用户群组设置
Figure PCTCN2022092996-appb-000001
共三个召回方案,
Figure PCTCN2022092996-appb-000002
的概率为50%、
Figure PCTCN2022092996-appb-000003
的概率为25%、
Figure PCTCN2022092996-appb-000004
的概率为25%,
Figure PCTCN2022092996-appb-000005
的子范围为[0,0.5)、
Figure PCTCN2022092996-appb-000006
的子范围为[0.5,0.75)、
Figure PCTCN2022092996-appb-000007
的子范围为[0.75,1],在总范围[0,1]之间随机生成一浮点数,如果浮点数位于[0,0.5)内,则选定
Figure PCTCN2022092996-appb-000008
如果浮点数位于[0.5,0.75),则选定
Figure PCTCN2022092996-appb-000009
如果浮点数位于[0.75,1]内,则选定
Figure PCTCN2022092996-appb-000010
S103、为第一用户执行召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户。
在多媒体平台中,在召回的流程中通常会配置不同的源,对于召回第二用户的场景,这些源可以称之为用户池。
这些用户池有不同的功能,可能对于不同的用户群组带来的收益有一定的差异性。
示例性地,用户池包括如下至少一种第二用户:
1、基于被第一用户关注的难易程度筛选的第二用户
在本示例中,可以召回容易被关注的第二用户,能够使得第一用户的关系链行为有提升,例如,召回的第二用户使得出现有效关系链时间的频次增多。
对于直播的业务的场景,可以对最近一段时间(如90天)内的,有一定比例(如15天以上)直播的时间大于预设的时间阈值(如10分钟)的第二用户,利用关注人数除以总曝光人数,得到曝光-关注转化率,将第二用户按照曝光-转化率对第二用户进行排序,取出一定比例(如前5%)的第二用户放入用户池。
2、基于多媒体数据被消费的高低程度筛选的第二用户
在本示例中,召回高被消费的第二用户,能够使得消费的时间增加,增加触发有效消费行为的第一用户。
对于直播的业务的场景,可以对最近一段时间(如90天)内的,有一定比例(如15天以上)直播的时间大于预设的时间阈值(如10分钟)的第二用户,利用第一用户平均观看的时长除以开播的时长,得到第一用户的平均观看时长占比,按照平均观看时长占比对第二用户进行排序,取出一定比例(如前5%)的第二用户放入用户池。
3、与第一用户个性化匹配的第二用户
在本示例中,可以应用神经网络、机器学习等训练个性化推荐模型,可以召回这些个性化模型针对第一用户个性化匹配的第二用户。
4、基于实时热度筛选的第二用户
在本示例中,对于直播等要求在线的业务场景,可以将第二用户在线的指标(如在线人数、平均观看时间等)通过加权求和等方式映射为热度,按照热度对第二用户进行排序,取出一定比例(如前5%)的第二用户放入用户池,在上述用户池中的第二用户离线时、保证召回第二用户的数量。
上述用户池只是作为示例,在实施本申请实施例时,可以根据实际情况设置其它用户池,本申请实施例对此不加以限制。
在一种召回方案的实现方式中,召回方案可以包括多个元素,元素的位置表示召回第二用户的用户池(即元素的位置与用户池(以ID等数据标识)存在映射关系),元素的数值表示从该用户池中召回第二用户的数量。
在本实施方式中,可以为第一用户在召回方案中读取为每个用户池中配置的元素,该元素具有位置、以及数值,从而为第一用户从位置对应的用户池中,召回数量为数值的第二用户。
例如,召回方案为
Figure PCTCN2022092996-appb-000011
其中,
Figure PCTCN2022092996-appb-000012
共有4个元素,第一个元素对应用户池A、第二个元素对应用户池B、第三个元素对应用户池C、第四个元素对应用户池D,在召回时,从用户池A中召回4个第二用户,从用户池B中召回5个第二用户,从用户池C召回9个第二用户,从用户池D中召回12个第二用户。
S104、将召回的第二用户的用户信息推送至客户端进行显示。
在实现中,如图2所示,推送第二用户的用户信息的过程通常包括如下操作:
1、召回
从用户池中召回第二用户,缩小可选的视频数据的集合。
2、粗排
召回的第二用户的数量较多,通常达到万、千这个量级,而精排使用的算法可能较为复杂,为了提高排序的速度,可以在召回和精排之间加入一个粗排的环节,通过少量第二用户、多媒体数据的特征,将召回的第二用户的数量加载至简单的排序模型中,例如,逻辑回归(Logistic Regression,LR)模型、梯度提升树(Gradient Boost Decision Tree,GBDT)模型,等等,对召回的第二用户进行粗略的排序,选择排序较高的部分第二用户,在保证一定精准的前提下,减少精排的第二用户的数量,一般可将第二用户的数量降至千、百这个量级。
根据业务场景的特性,粗排往往是可选的,即可以应用粗排,也可以直接从召回跳转到精排,本实施例对此不加以限制。
3、精排
通过较多的第二用户、多媒体数据的特征,将粗排的第二用户的数量加载至较为复杂的排序模型中,例如,卷积神经网络(Convolutional Neural Networks,CNN)、循环神经网络(Recurrent Neural Network,RNN),等等,对粗排的视频数据进行精确的排序,选择排序较高的部分第二用户,尽量提高排序的精准度,减少发送至客户端的第二用户的数量,一般可将第二用户的数量降至百、 十这个量级。
在本实施例中,对于精排之后所提取的第二用户,可以称之为用户序列,在打散(又称之为重排)用户序列中的第二用户之后,维持第二用户的数量(如百、十),将该第二用户的用户信息(如名称、头像等)发送至客户端,客户端按照顺序显示多个第二用户的用户信息,供第一用户浏览。
若第一用户发现感兴趣的第二用户,则可以通过鼠标点击、触控点击等方式触发该第二用户的用户信息,客户端向多媒体平台请求该第二用户的页面,例如,第二用户正在主持直播节目的页面、聚合第二用户所发布的多媒体数据的页面,等等,加载该页面供第二用户浏览。
在本实施例中,当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组,用户群组关联多个配置有权重的召回方案,以用户群组而非单个用户作为优化召回方案的目标,可提高用户对相同召回方案反馈的数据量,可以降低单个用户对召回方案反馈的数据量少导致不置信的情况,即使单个用户反馈的数据有波动或者噪音也是可以接受的,根据多个召回方案分别对应的多个权重为第一用户选择召回方案,为第一用户执行召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户,将召回的第二用户的用户信息推送至客户端进行显示,考虑到不同用户池在内容质量、用户偏好和指标贡献等方面具有差异性,从不同用户池召回第二用户、使得不同用户群组能够有相适应的配额,提高了召回的多样性,可提高召回第一用户感兴趣的第二用户的情况,也便于协助后期对召回方案的迭代,优化召回方案。
第一用户可以直接触发感兴趣的第二用户的用户信息,无需通过关键词搜索感兴趣的第二用户,大大提高了获取感兴趣信息的简便性,减少对多媒体平台召回第二用户、推送第二用户的用户信息、第一用户的设备显示第二用户的用户信息的资源。
实施例二
图3为本申请实施例二提供的一种用户召回方法的流程图,本实施例以前述实施例为基础,增加更新召回方案的操作,该方法包括:
S301、当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组。
如图4所示,用户群组关联多个配置有权重的召回方案,用户群组中包括多个第一用户,考虑到实时更新召回方案,不使用第一用户的历史数据,因此,用户群组中的第一用户小幅度变更时,一般不重新初始化用户群组的召回方案, 如果用户群组产生较大幅度的变更,例如,新增用户群组、删除用户群组,等等,则可以初始化用户群组的召回方案。
将用户群组记为G,确定用户群组G中召回方案的数量上限N,以及每个召回方案中召回第二用户的总数M。
对用户群组G配置N个召回方案
Figure PCTCN2022092996-appb-000013
到权重W的映射所构成的集合,即
Figure PCTCN2022092996-appb-000014
保证
Figure PCTCN2022092996-appb-000015
(即,召回方案
Figure PCTCN2022092996-appb-000016
中多个元素的数值之和为总数M),初始化时,召回方案中多个元素的数值相同,所有召回方案的权重相同且和为1。
S302、根据多个召回方案分别对应的多个权重为第一用户选择召回方案。
S303、为第一用户执行召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户。
在本实施例中,技术人员可以通过优化召回方案的结果来对用户池进行更新,如新增用户池、删除用户池等,使得整个召回的过程更加高效,协助召回方案的迭代。
如图4所示,当更新用户池时,使用该用户池进行召回的每个召回方案随之进行更新。
召回方案包括多个元素,元素的位置表示召回第二用户的用户池,元素的数值表示召回第二用户的数量。
当新增用户池时,召回方案的长度增加一,在召回方案中新增一个指示该用户池的元素,记录新增的元素的位置与新增的用户池之间的映射关系,将新增的元素的数值设置为其他元素的数值的平均值,若限定多个召回方案中召回第二用户的总数,则可以按比例对当前多个元素的数值进行缩放、并进行取整操作(包括向上取整、向下取整),使得召回方案中多个元素的数值之和为该总数。
当删除用户池时,召回方案的长度减少一,在召回方案中删除指示该用户池的元素,即,按照映射关系查找待删除的用户池映射的位置,删除位于该位置的元素,若限定多个召回方案中召回第二用户的总数,则可以按比例对当前多个元素的数值进行缩放、并进行取整操作(包括向上取整、向下取整),使得召回方案中多个元素的数值之和为该总数。
S304、将召回的第二用户的用户信息推送至客户端进行显示。
如图4所示,执行召回方案,按照召回方案的配置,从相应的用户池中召回第二用户,并将第二用户的用户信息推送至客户端、供第一用户浏览。
S305、统计第一用户在每个召回方案下产生的、用于评价第二用户的指标。
本实施例推送第二用户的用户信息是实时的,因此,评价第二用户的指标是具有一定的实时性。由于采用了用户群组而非单个用户作为优化的目标,因此,即使单个用户评价第二用户的指标有波动或者噪音也是可以接受的。
如图4所示,按照优化的业务场景、目标、获取指标的难度程度、获取指标的实时性强弱等因素,可以设置不同的、用于评价第二用户的指标,例如,指标为第一用户点击第二用户的用户信息的情况(如点击量、点击率),消费第二用户生产的多媒体数据的时长,第一用户是否关注第二用户,等等。
这些指标可以单个使用,也可以组合使用、生成最终的指标,例如,对多个指标进行加权平均,对多个指标归一化之后相乘,等等,本实施例对此不加以限制。
在本实施例中,通过以特定规则划分第一用户的用户群组为对象,利用相适应的规则评价第二用户的指标,可以逐步优化召回方案,能够使得用户群组按照所希望的方向进行发展,优化用户的体验,提高用户的渗透,增加多媒体平台的收益。
S306、根据指标更新每个召回方案的权重。
如图4所示,第一用户在每个召回方案下产生的指标,可以在一定程度上反映了召回方案的质量,可以以此为依据,对召回方案的权重进行更新。
一般情况下,第一用户反馈更加正向的指标,表示召回方案的质量更优,此时可以上调该召回方案的权重,第一用户反馈更加负向的指标,表示召回方案的质量更差,此时可以下降该召回方案的权重。
在一种更新的方式中,由于每个召回方案的权重不同,导致下发的第一用户的数量或获得的指标的数量会有差异,如果一个召回方案下发的第一用户的数量较少,可能会因为波动导致指标不置信。
考虑到上述原因,在本方式中,可衡量召回方案的相对优劣,而无需估计召回方案的期望指标大小,因此,可对所有召回方案按照指标的大小进行排序,按照一定的比例将多个指标划分为第一候选指标、以及第二候选指标,其中,第一候选指标高于第二候选指标,以50%为例,第一候选指标为数值最高的50%个指标,第二候选指标为数值最低的50%个指标。
将第一候选指标作为正例、将第二候选指标作为负例,统计正例的数量、负例的数量,使用威尔逊区间法、正态区间(Normal approximation interval)等方式对召回方案计算置信区间,从而基于置信区间修正权重,将修正后的权重作为召回方案的新的权重。
在一种修正的方式中,可以使用平滑的方式修正权重。
一方面,可计算置信区间的下限值与第一系数之间的第一乘积。
置信区间的实质,就是进行可信度的修正,弥补样本量(即第一用户反馈的指标)较少的影响。如果样本量多,就说明比较可信,进行较少的修正,此时置信区间会比较窄,下限值会比较大;如果样本量少,就说明不一定可信,进行较大的修正,此时置信区间会比较宽,下限值会比较小。
以威尔逊区间法为例,置信区间的下限值如下计算:
n=u+v
Figure PCTCN2022092996-appb-000017
Figure PCTCN2022092996-appb-000018
其中,u为正例、v为负例,z α是与置信区间相关的常数,可以取2,Score为置信区间的下限值。
另一方面,可计算权重与第二系数之间的第二乘积,其中,第二系数大于第一系数,可以应对偶然的波动。
计算第一乘积与第二乘积之间的和值,作为召回方案新的权重,那么,权重的更新过程表示如下:
W=αW new+βW old
其中,W为更新之后的权重,W new为置信区间的下限值,W old为原始的权重,α为第二系数,如0.25,β为第一系数,如0.75。
S307、使用更新之后的权重在召回方案的基础上为用户群组生成新的召回方案、并为新的召回方案配置新的权重。
对于更新权重之后的召回方案,可以筛选出优质的召回方案,以此为基础生成新的召回方案,同时,以该召回方案更新之后的权重为基础,为新的召回方案配置新的权重,此优化过程不断迭代,可以筛选出更加优质的最优解。
在本申请的一个实施例中,对于更新权重之后的召回方案,可以使用遗传算法(Genetic Algorithm,GA)对召回方案及其权重进行遗传操作,遗传较为优质的召回方案的基因,从而生成新的召回方案、为新的召回方案配置新的权重。
遗传算法以一种种群(同一个用户群组的召回方案)中的所有个体(召回方案)为对象,并利用随机化技术指导对一个被编码的参数空间进行高效搜索。 其中,选择、交叉和变异构成了遗传算法的遗传操作。
初代种群(同一个用户群组的召回方案)产生之后,按照适者生存和优胜劣汰的原理,逐代(generation)演化产生出越来越好的近似解(同一个用户群组的召回方案),在每一代种群中,根据问题域(评价第二用户的指标)中个体的适应度(fitness)大小选择(selection)个体,并借助于自然遗传学的遗传算子(genetic operators)进行组合交叉(crossover)和变异(mutation),产生出代表新的解集的种群(同一个用户群组的召回方案)。
这个过程将导致种群像自然进化一样的后生代种群比前代更加适应于环境,末代种群中的最优个体经过解码(decoding),可以作为问题近似最优解。
在本实施例中,S307包括:
S3071、将多个召回方案划分为第一目标方案、第二目标方案。
S3072、删除第二目标方案。
如图4所示,在本实施例中,可对召回方案进行选择操作,即,对所有召回方案按照权重的大小进行排序,按照一定的比例将多个召回方案划分为第一目标方案、以及第二目标方案,使得第一目标方案的权重高于第二目标方案的权重,以50%为例,第一目标方案为权重最高的50%个召回方案,第二目标方案为数值最低的50%个召回方案。
删除第二目标方案,保留更优的第一目标方案进行交叉操作、变异操作。
S3073、在第一目标方案之间交换部分元素、为用户群组生成新的召回方案,基于第一目标方案的权重为新的召回方案配置新的权重。
如图4所示,在本实施例中,可对部分第一目标方案进行交叉操作,在交叉操作中,可以选择任意第一目标方案之间交换部分元素,交互元素之后的第一目标方案为用户群组新的召回方案,参考参与交叉操作的第一目标方案的权重为新的召回方案配置新的权重。
在一种交叉操作中,可按照权重对参与交叉操作的第一目标方案进行排序,遍历已排序的第一目标方案,针对当前第一目标方案,搜索权重高于当前第一目标方案的其他第一目标方案,以便在交叉操作中遗传更优质的基因,提高新的召回方案的质量。
例如,一个用户群组共有
Figure PCTCN2022092996-appb-000019
(权重为0.4)、
Figure PCTCN2022092996-appb-000020
(权重为0.3)、
Figure PCTCN2022092996-appb-000021
(权重为0.2)、
Figure PCTCN2022092996-appb-000022
(权重为0.1)共四个第一目标方案,则可以在
Figure PCTCN2022092996-appb-000023
Figure PCTCN2022092996-appb-000024
之间进行交叉操作,在
Figure PCTCN2022092996-appb-000025
Figure PCTCN2022092996-appb-000026
Figure PCTCN2022092996-appb-000027
之间进行交叉操作,在
Figure PCTCN2022092996-appb-000028
Figure PCTCN2022092996-appb-000029
Figure PCTCN2022092996-appb-000030
Figure PCTCN2022092996-appb-000031
之间进行交叉操作。
将当前第一目标方案与其他第一目标方案中处于相同位置的部分(如20%)元素的数值互相交换,作为用户群组新的召回方案。
例如,假设第一目标方案
Figure PCTCN2022092996-appb-000032
Figure PCTCN2022092996-appb-000033
之间交互位于第3位、第4为元素的数值,生成新的召回方案
Figure PCTCN2022092996-appb-000034
或者,
Figure PCTCN2022092996-appb-000035
若限定多个召回方案中召回第二用户的总数,基于交叉操作生成的新的召回方案中多个元素的数值之和并不一定等于该总数,此时,可以按比例对多个元素的数值进行缩放、并进行取整操作(包括向上取整、向下取整),使得召回方案中多个元素的数值之和为该总数。
此外,考虑到新的召回方案的质量存在不确定性,可减少新的召回方案的曝光机会,将参与交叉操作的第一目标方案中的权重的较小者(即当前第一目标方案的权重)赋值至新的召回方案的权重,若新的召回方案的质量较好,后续可以在更新权重时,提高其权重,若新的召回方案的质量较差,较少的曝光机会可以减少对推送第二用户的用户信息的效果的影响。
S3074、变更第一目标方案的部分元素、为用户群组生成新的召回方案,基于第一目标方案的权重为新的召回方案配置新的权重。
如图4所示,在本实施例中,可对部分第一目标方案进行变异操作,在变异操作中,可以选择任意第一目标方案变更部分元素,变更元素之后的第一目标方案为用户群组新的召回方案,参考参与变更操作的第一目标方案的权重为新的召回方案配置新的权重。
一般情况下,参与了交叉操作的第一目标方案,并不参与变异操作,如果限定了同一个用户群组的召回方案的数量N,可以持续对第一目标方案进行变异操作,直到所有的召回方案的数量达到上限N,如果所有的第一目标方案执行变异操作仍未达到上限N,则可以将在前一次变异操作中生成的召回方案作为新的第一目标操作再次执行变异操作,直至所有的召回方案的数量达到上限N。
在一种变异操作中,可按照权重对参与变异操作的第一目标方案进行排序,遍历已排序的第一目标方案,针对当前第一目标方案,基于第一目标方案的元素的数值确定数值区间,数值区间的下限值可以大于、小于或等于第一目标方案的元素中最小的数值,数值区间的上限值可以大于、小于或等于第一目标方案的元素中最大的数值,示例性地,将第一目标方案的元素中最小的数值减一作为数值区间的下限值,将第一目标方案的元素中最大的数值加一作为数值区间的上限值。
通过随机等方式随机在第一目标方案中选定部分(如20%)元素,针对第 一目标方案的该部分元素,随机在数值区间中取值、以该取值替换该部分元素的数值。
若限定多个召回方案中召回第二用户的总数,基于变异操作生成的新的召回方案中多个元素的数值之和并不一定等于该总数,此时,可以按比例对多个元素的数值进行缩放、并进行取整操作(包括向上取整、向下取整),使得召回方案中多个元素的数值之和为该总数。
此外,考虑到新的召回方案的质量存在不确定性,可减少新的召回方案的曝光机会,对第一目标方案的权重取指定的比例(该比例小于1,如80%),作为新的召回方案的权重,若新的召回方案的质量较好,后续可以在更新权重时,提高其权重,若新的召回方案的质量较差,较少的曝光机会可以减少对推送第二用户的用户信息的效果的影响。
若限定同一个用户群组的所有召回方案的权重的和为1,此时,对用户群组所有的召回方案的权重按照比例进行归一化处理,使得该用户群组所有的召回方案的权重之和为1。
上述遗传操作只是作为优化召回方案的操作的示例,在实施本申请实施例时,可以根据实际情况设置其它优化召回方案的操作,例如,粒子群算法、蚂蚁算法、梯度下降法,等等,本申请实施例对此不加以限制。
此外,进行其他实验时,为了控制变量,可能关闭上述实时优化召回方案的操作,此时,对于每个用户群组,可选择权重最高的召回方案作为最终的召回方案,重新开启上述实时优化召回方案的操作时,可以跳过初始化,通过连续的变异操作使得一个用户群组的召回方案的数量达到上限N,然后,按流程优化召回方案。
在本实施例中,统计第一用户在每个召回方案下产生的、用于评价第二用户的指标,根据指标更新每个召回方案的权重,使用更新之后的权重在召回方案的基础上为用户群组生成新的召回方案、并为新的召回方案配置新的权重,由于评价第二用户的指标可以实时获取,且更新权重、迭代优化的操作较为简单,因此,对召回方案的优化可以保证实时性,大大提高了召回方案与第一用户之间的适配度,相比使用固定的召回方案,本实施例根据第一用户对召回方案的反馈,对召回方案的权重进行更新,从而作用于召回方案的优化,可提高召回的效果,避免了人为设计导致的召回效果不理想的情况,对召回方案进行优化,较优的用户池有更大的发挥空间,较差的用户池会逐步被淘汰掉,使可提高召回方案的质量。
对于方法实施例,为了简单描述,故将其都表述为一系列的动作组合,但是,本申请实施例并不受所描述的动作顺序的限制,因为依据本申请实施例,一些步骤可以采用其他顺序或者同时进行。其次,文中所描述的实施例所涉及的动作并不一定是本申请实施例所必须的。
实施例三
图5为本申请实施例三提供的一种用户召回装置的结构框图,可以包括如下模块:
用户群组确定模块501,设置为当接收到客户端发送的请求时,确定用于消费多媒体数据的第一用户所处的用户群组,所述用户群组关联多个配置有权重的召回方案;召回方案选择模块502,设置为根据多个召回方案分别对应的多个权重为所述第一用户选择召回方案;召回方案执行模块503,设置为为所述第一用户执行所述召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户;用户信息推送模块504,设置为将召回的第二用户的用户信息推送至所述客户端进行显示。
本申请实施例所提供的用户召回装置可执行本申请任意实施例所提供的用户召回方法,具备执行方法相应的功能模块和效果。
实施例四
图6为本申请实施例四提供的一种计算机设备的结构示意图。图6示出了适于用来实现本申请实施方式的示例性计算机设备12的框图。
计算机设备12的组件可以包括但不限于:一个或者多个处理器或者处理单元16,内存28,连接不同系统组件(包括内存28和处理单元16)的总线18。
计算机设备12包括多种计算机系统可读介质。这些介质可以是任何能够被计算机设备12访问的可用介质,包括易失性和非易失性介质,可移动的和不可移动的介质。
内存28可以包括易失性存储器形式的计算机系统可读介质,例如随机存取存储器(Random Access Memory,RAM)30和/或高速缓存32。存储系统34可以用于读写不可移动的、非易失性磁介质(图6未显示,通常称为“硬盘驱动器”)。内存28可以包括至少一个程序产品,该程序产品具有一组(例如至少一个)程序模块,这些程序模块被配置以执行本申请实施例的功能。
具有一组(至少一个)程序模块42的程序/实用工具40,可以存储在例如 内存28中。程序模块42通常执行本申请所描述的实施例中的功能和/或方法。
计算机设备12也可以与一个或多个外部设备14(例如键盘、指向设备、显示器24等)通信。这种通信可以通过输入/输出(I/O)接口22进行。并且,计算机设备12还可以通过网络适配器20与一个或者多个网络(例如局域网(Local Area Network,LAN),广域网(Wide Area Network,WAN)和/或公共网络。
处理单元16通过运行存储在内存28中的程序,从而执行多种功能应用以及数据处理,例如实现本申请实施例所提供的用户召回方法。
实施例五
本申请实施例五还提供一种计算机可读存储介质,计算机可读存储介质上存储有计算机程序,该计算机程序被处理器执行时实现上述用户召回方法的多个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
在本文中,计算机可读存储介质可以是任何包含或存储程序的有形介质,该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。

Claims (17)

  1. 一种用户召回方法,包括:
    在接收到客户端发送的请求的情况下,确定用于消费多媒体数据的第一用户所处的用户群组,其中,所述用户群组关联多个配置有权重的召回方案;
    根据多个召回方案分别对应的多个权重为所述第一用户选择召回方案;
    为所述第一用户执行所述召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户;
    将召回的第二用户的用户信息推送至所述客户端进行显示。
  2. 根据权利要求1所述的方法,其中,所述确定用于消费多媒体数据的第一用户所处的用户群组,包括:
    统计第一用户在每个第一时间段内消费多媒体数据所产生的行为数据,其中,第二时间段包括多个第一时间段;
    基于所述行为数据识别所述每个第一时间段的属性;
    在所述多个第一时间段的属性在所述第二时间段内符合预设的群组条件的情况下,将所述第一用户划分至所述群组条件对应的用户群组。
  3. 根据权利要求2所述的方法,其中,所述基于所述行为数据识别所述每个第一时间段的属性,包括:
    在所述第一用户在所述每个第一时间段内消费多媒体数据的总时间大于第一时间阈值的情况下,对所述第一时间段标记属性为有效时间;
    在所述第一用户在所述第一时间段内消费目标数据的总时间大于或等于第一时间阈值的情况下,对所述第一时间段标记属性为有效关系链时间;
    其中,所述目标数据为所述第一用户关注的第二用户生产的多媒体数据。
  4. 根据权利要求3所述的方法,其中,所述在所述多个第一时间段的属性在所述第二时间段内符合预设的群组条件的情况下,将所述第一用户划分至所述群组条件对应的用户群组,包括:
    统计所述有效时间在所述第二时间段内出现的第一频次、并统计所述有效关系链时间在所述第二时间段内出现的第二频次;
    计算所述第二频次与所述第一频次之间的比值;
    在所述第一频次小于第一频次阈值的情况下,确定所述第一用户归属的用户群组为新用户群组或低活跃群组;
    在所述第一频次大于第一频次阈值、且小于第二频次阈值,并且,所述比 值小于第一比例阈值的情况下,确定所述第一用户归属的用户群组为低关系链低有效群组;
    在所述第一频次大于第一频次阈值、且小于第二频次阈值,并且,所述比值大于第一比例阈值的情况下,确定所述第一用户归属的用户群组为低关系链高有效群组;
    在所述第一频次大于第二频次阈值,并且,所述比值小于第一比例阈值的情况下,确定所述第一用户归属的用户群组为高有效低关系链群组;
    在所述第一频次大于第二频次阈值,并且,所述比值大于第一比例阈值的情况下,则确定所述第一用户归属的用户群组为高有效高关系链群组。
  5. 根据权利要求1所述的方法,其中,所述根据多个召回方案分别对应的多个权重为所述第一用户选择召回方案,包括:
    将每个召回方案的权重映射为概率,其中,所述概率与所述权重正相关;
    根据所述多个召回方案分别对应的多个概率为所述第一用户选择所述多个召回方案中的一个召回方案。
  6. 根据权利要求1所述的方法,其中,所述为所述第一用户执行所述召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户,包括:
    为所述第一用户在所述召回方案中读取为每个用户池中配置的元素,其中,所述元素具有位置以及数值;
    为所述第一用户从所述位置对应的用户池中,召回数量为所述数值的第二用户。
  7. 根据权利要求1-6中任一项所述的方法,其中,所述用户池包括如下至少一种第二用户:
    基于被所述第一用户关注的难易程度筛选的第二用户;
    基于多媒体数据被消费的高低程度筛选的第二用户;
    与所述第一用户个性化匹配的第二用户;
    基于实时热度筛选的第二用户。
  8. 根据权利要求1-6中任一项所述的方法,还包括:
    统计所述第一用户在每个召回方案下产生的、用于评价所述第二用户的指标;
    根据所述指标更新所述每个召回方案的权重;
    使用更新之后的权重在所述召回方案的基础上为所述用户群组生成新的召回方案、并为所述新的召回方案配置新的权重。
  9. 根据权利要求8所述的方法,其中,在所述指标的数量为多个的情况下,所述根据所述指标更新所述每个召回方案的权重,包括:
    将多个指标划分为第一候选指标、以及第二候选指标,其中,所述第一候选指标高于所述第二候选指标;
    将所述第一候选指标作为正例,并将所述第二候选指标作为负例,对所述召回方案计算置信区间;
    基于所述置信区间修正所述权重,将修正后的权重作为所述召回方案的新的权重。
  10. 根据权利要求9所述的方法,其中,所述基于所述置信区间修正所述权重,将修正后的权重作为所述召回方案的新的权重,包括:
    计算所述置信区间的下限值与第一系数之间的第一乘积;
    计算所述权重与第二系数之间的第二乘积,其中,所述第二系数大于所述第一系数;
    计算所述第一乘积与所述第二乘积之间的和值,将所述和值作为所述召回方案的新的权重。
  11. 根据权利要求8所述的方法,其中,所述使用更新之后的权重在所述召回方案的基础上为所述用户群组生成新的召回方案、并为所述新的召回方案配置新的权重,包括:
    将所述多个召回方案划分为第一目标方案、以及第二目标方案,其中,所述第一目标方案的权重高于所述第二目标方案的权重;
    删除所述第二目标方案;
    在所述第一目标方案之间交换部分元素、为所述用户群组生成新的召回方案,基于所述第一目标方案的权重为所述新的召回方案配置新的权重;
    变更所述第一目标方案的部分元素、为所述用户群组生成新的召回方案,基于所述第一目标方案的权重为所述新的召回方案配置新的权重。
  12. 根据权利要求11所述的方法,其中,所述在所述第一目标方案之间交换部分元素、为所述用户群组生成新的召回方案,基于所述第一目标方案的权重为所述新的召回方案配置新的权重,包括:
    针对当前第一目标方案中,搜索权重高于所述当前第一目标方案的其他第 一目标方案;
    将所述当前第一目标方案与所述其他第一目标方案中处于相同位置的部分元素的数值互相交换,作为所述用户群组新的召回方案;
    将所述当前第一目标方案的权重赋值至所述新的召回方案的权重。
  13. 根据权利要求11所述的方法,其中,所述变更所述第一目标方案的部分元素、为所述用户群组生成新的召回方案,基于所述第一目标方案的权重为所述新的召回方案配置新的权重,包括:
    基于所述第一目标方案的元素的数值确定数值区间;
    针对所述第一目标方案的部分元素,随机在所述数值区间中取值、以所述取值替换所述部分元素的数值;
    对所述第一目标方案的权重取指定的比例,将取指定的比例后的权重作为所述新的召回方案的权重。
  14. 根据权利要求1-6中任一项所述的方法,还包括:
    在新增用户池的情况下,在所述召回方案中增加指示所述用户池的元素,将新增的元素中的数值设置为所述召回方案中除所述新增的元素外的元素的数值的平均值;
    在删除用户池的情况下,在所述召回方案中删除指示所述用户池的元素。
  15. 一种用户召回装置,包括:
    用户群组确定模块,设置为在接收到客户端发送的请求的情况下,确定用于消费多媒体数据的第一用户所处的用户群组,其中,所述用户群组关联多个配置有权重的召回方案;
    召回方案选择模块,设置为根据多个召回方案分别对应的多个权重为所述第一用户选择召回方案;
    召回方案执行模块,设置为为所述第一用户执行所述召回方案、从多个用户池中分别召回用于生产多媒体数据的第二用户;
    用户信息推送模块,设置为将召回的第二用户的用户信息推送至所述客户端进行显示。
  16. 一种计算机设备,包括:
    至少一个处理器;
    存储器,设置为存储至少一个程序;
    当所述至少一个程序被所述至少一个处理器执行,使得所述至少一个处理器实现如权利要求1-14中任一项所述的用户召回方法。
  17. 一种计算机可读存储介质,所述计算机可读存储介质上存储计算机程序,所述计算机程序被处理器执行时实现如权利要求1-14中任一项所述的用户召回方法。
PCT/CN2022/092996 2021-05-24 2022-05-16 用户召回方法、装置、计算机设备和存储介质 WO2022247671A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110563301.5A CN113297398B (zh) 2021-05-24 2021-05-24 一种用户召回方法、装置、计算机设备和存储介质
CN202110563301.5 2021-05-24

Publications (1)

Publication Number Publication Date
WO2022247671A1 true WO2022247671A1 (zh) 2022-12-01

Family

ID=77324202

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/092996 WO2022247671A1 (zh) 2021-05-24 2022-05-16 用户召回方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN113297398B (zh)
WO (1) WO2022247671A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113297398B (zh) * 2021-05-24 2024-06-21 百果园技术(新加坡)有限公司 一种用户召回方法、装置、计算机设备和存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110096646A (zh) * 2019-05-08 2019-08-06 广州虎牙信息科技有限公司 品类关联信息的生成及其视频推送方法和相关设备
CN110585726A (zh) * 2019-09-16 2019-12-20 腾讯科技(深圳)有限公司 用户召回方法、装置、服务器及计算机可读存储介质
CN112115370A (zh) * 2020-09-29 2020-12-22 贝壳技术有限公司 推荐方法、装置、计算机可读存储介质及电子设备
CN112528164A (zh) * 2020-12-14 2021-03-19 建信金融科技有限责任公司 一种用户协同过滤召回方法及装置
US20210097133A1 (en) * 2019-09-27 2021-04-01 Microsoft Technology Licensing, Llc Personalized proactive pane pop-up
CN112749329A (zh) * 2020-04-29 2021-05-04 腾讯科技(深圳)有限公司 内容搜索方法、装置、计算机设备及存储介质
CN113297398A (zh) * 2021-05-24 2021-08-24 百果园技术(新加坡)有限公司 一种用户召回方法、装置、计算机设备和存储介质

Family Cites Families (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030164844A1 (en) * 2000-09-25 2003-09-04 Kravitz Dean Todd System and method for processing multimedia content, stored in a computer-accessible storage medium, based on various user-specified parameters related to the content
US8548940B1 (en) * 2008-09-12 2013-10-01 Salesforce.Com, Inc. System, method and computer program product for executing recall actions with respect to an approval process in a multi-tenant an on-demand database service
US20110276394A1 (en) * 2010-05-05 2011-11-10 Positioniq, Inc. Automated Targeted Information System
JP6416842B2 (ja) * 2016-10-03 2018-10-31 日本瓦斯株式会社 リコール対象機器判定システム
US10715848B2 (en) * 2018-05-09 2020-07-14 Pluto Inc. Methods and systems for generating and providing program guides and content
CN112395489B (zh) * 2019-08-15 2023-04-11 中移(苏州)软件技术有限公司 一种推荐方法及装置、设备和计算机存储介质
CN112711945B (zh) * 2019-10-25 2022-08-19 上海哔哩哔哩科技有限公司 广告召回方法和系统
CN111008278B (zh) * 2019-11-22 2022-06-21 厦门美柚股份有限公司 内容推荐方法及装置
CN110990695A (zh) * 2019-11-22 2020-04-10 厦门美柚股份有限公司 推荐系统内容召回方法及装置
CN110996116B (zh) * 2019-12-18 2022-03-11 广州市百果园信息技术有限公司 一种主播信息的推送方法、装置、计算机设备和存储介质
CN111178970B (zh) * 2019-12-30 2023-06-30 微梦创科网络科技(中国)有限公司 广告投放的方法及装置、电子设备和计算机可读存储介质
CN111918104A (zh) * 2020-07-29 2020-11-10 有半岛(北京)信息科技有限公司 一种视频数据的召回方法、装置、计算机设备和存储介质
CN112116393B (zh) * 2020-09-23 2022-03-15 贝壳找房(北京)科技有限公司 用于实现事件用户维护的方法、装置和设备
CN112287167A (zh) * 2020-10-29 2021-01-29 四川长虹电器股份有限公司 视频推荐召回方法及装置
CN112765241B (zh) * 2021-02-04 2024-06-11 腾讯科技(深圳)有限公司 召回数据确定方法、装置及存储介质

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110096646A (zh) * 2019-05-08 2019-08-06 广州虎牙信息科技有限公司 品类关联信息的生成及其视频推送方法和相关设备
CN110585726A (zh) * 2019-09-16 2019-12-20 腾讯科技(深圳)有限公司 用户召回方法、装置、服务器及计算机可读存储介质
US20210097133A1 (en) * 2019-09-27 2021-04-01 Microsoft Technology Licensing, Llc Personalized proactive pane pop-up
CN112749329A (zh) * 2020-04-29 2021-05-04 腾讯科技(深圳)有限公司 内容搜索方法、装置、计算机设备及存储介质
CN112115370A (zh) * 2020-09-29 2020-12-22 贝壳技术有限公司 推荐方法、装置、计算机可读存储介质及电子设备
CN112528164A (zh) * 2020-12-14 2021-03-19 建信金融科技有限责任公司 一种用户协同过滤召回方法及装置
CN113297398A (zh) * 2021-05-24 2021-08-24 百果园技术(新加坡)有限公司 一种用户召回方法、装置、计算机设备和存储介质

Also Published As

Publication number Publication date
CN113297398B (zh) 2024-06-21
CN113297398A (zh) 2021-08-24

Similar Documents

Publication Publication Date Title
CN110929052B (zh) 多媒体资源推荐方法、装置、电子设备及存储介质
CN110781321B (zh) 一种多媒体内容推荐方法及装置
CN109033408B (zh) 信息推送方法及装置、计算机可读存储介质、电子设备
WO2021135562A1 (zh) 特征有效性评估方法、装置、电子设备及存储介质
CN111126495B (zh) 模型训练方法、信息预测方法、装置、存储介质及设备
WO2016107354A1 (zh) 提供用户个性化资源消息推送的方法和装置
CN112508609A (zh) 人群扩量的预测方法、装置、设备及存储介质
CN113015010A (zh) 推送参数确定方法、装置、设备及计算机可读存储介质
WO2020258773A1 (zh) 确定推送用户群的方法、装置、设备及存储介质
CN111695084A (zh) 模型生成方法、信用评分生成方法、装置、设备及存储介质
CN112131413A (zh) 一种多媒体信息处理方法、装置、电子设备及存储介质
CN114417058A (zh) 一种视频素材的筛选方法、装置、计算机设备和存储介质
WO2022247671A1 (zh) 用户召回方法、装置、计算机设备和存储介质
CN112884529A (zh) 一种广告竞价方法、装置、设备及介质
CN113656681A (zh) 一种对象评价方法、装置、设备及存储介质
CN112995690A (zh) 直播内容品类识别方法、装置、电子设备和可读存储介质
Doshi et al. Predicting movie prices through dynamic social network analysis
US20240202058A1 (en) Methods and systems for determining stopping point
WO2023087933A1 (zh) 内容推荐方法、装置、设备、存储介质及程序产品
CN116257758A (zh) 模型训练方法、人群拓展方法、介质、装置和计算设备
US11463461B2 (en) Unequal probability sampling based on a likelihood model score to evaluate prevalence of inappropriate entities
CN113450127A (zh) 信息展示方法、装置、计算机设备及存储介质
CN116861080A (zh) 信息推荐方法、装置、设备、存储介质以及产品
WO2024103620A1 (zh) 内容生成方法、装置、计算机设备和存储介质
CN113538030B (zh) 一种内容推送方法、装置及计算机存储介质

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: 22810398

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22810398

Country of ref document: EP

Kind code of ref document: A1