CN111881940B

CN111881940B - Live broadcast continuous wheat matching method and device, electronic equipment and storage medium

Info

Publication number: CN111881940B
Application number: CN202010609897.3A
Authority: CN
Inventors: 陈洋溢; 陈杰; 何楚; 彭伟湘
Original assignee: Guangzhou Cubesili Information Technology Co Ltd
Current assignee: Guangzhou Cubesili Information Technology Co Ltd
Priority date: 2020-06-29
Filing date: 2020-06-29
Publication date: 2023-11-24
Anticipated expiration: 2040-06-29
Also published as: CN111881940A

Abstract

The application discloses a live broadcast continuous wheat matching method, a device, electronic equipment and a storage medium, wherein the method comprises the following steps: acquiring a plurality of anchor to be matched in a matching pool; acquiring a comparison combination set among a plurality of anchor to be matched, wherein the comparison combination set consists of two different anchor to be matched in the plurality of anchor to be matched; obtaining a predicted total ratio of each ratio combination in a ratio combination set based on a pre-trained ratio prediction model, wherein the ratio prediction model is obtained by training a neural network model according to a training sample; and determining a plurality of comparison combination with the maximum sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values, and taking the comparison combination as a live-broadcast continuous-matching result of a plurality of main broadcasting to be matched, wherein the same main broadcasting to be matched does not exist in the comparison combination. According to the application, the anchor in the matching pool is matched by predicting the total comparison value, so that the live broadcast effect of the link wheat PK can be improved, and the conversion rate and the success rate of the live broadcast E-commerce are promoted.

Description

Live broadcast continuous wheat matching method and device, electronic equipment and storage medium

Technical Field

The application relates to the technical field of internet live broadcasting, in particular to a live broadcasting and wheat connecting matching method, a live broadcasting and wheat connecting matching device, electronic equipment and a storage medium.

Background

With the rapid development of internet technology, multimedia live broadcasting is widely focused by people in a novel form and rich in content, and with the gradual increase of users watching live broadcasting, the functions of a live broadcasting platform are also continuously developed. Wherein, the link wheat PK (spell) between the anchor is one of them. However, the matching mechanism of the existing link-microphone PK is single, and is usually random matching, so that the problem of large live broadcast level difference of the two parties as a main broadcast is easy to occur, and the live broadcast interaction effect is poor.

Disclosure of Invention

The embodiment of the application provides a live broadcast link wheat matching method, a live broadcast link wheat matching device, electronic equipment and a storage medium, which can improve the live broadcast effect of link wheat PK.

In a first aspect, an embodiment of the present application provides a live link matching method, where the method includes: acquiring a plurality of anchor to be matched in a matching pool; acquiring a comparison combination set among the plurality of anchor to be matched, wherein the comparison combination set comprises at least one comparison combination which is composed of two different anchor to be matched in the plurality of anchor to be matched; obtaining a predicted total comparison value of each comparison combination in the comparison combination set based on a pre-trained comparison prediction model, wherein the comparison prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a comparison combination sample and a total comparison value sample corresponding to the comparison combination sample; and determining a plurality of comparison combinations with the maximum sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values, wherein the comparison combinations are used as the matching results of the live-broadcast continuous-wheat comparison of the plurality of the hosts to be matched, and the same hosts to be matched do not exist in the comparison combinations.

In a second aspect, an embodiment of the present application provides a live-broadcast continuous-wheat matching device, where the device includes: the anchor acquisition module is used for acquiring a plurality of anchors to be matched in the matching pool; the combination acquisition module is used for acquiring a comparison combination set among the plurality of anchor to be matched, wherein the comparison combination set comprises at least one comparison combination which is formed by two different anchor to be matched in the plurality of anchor to be matched; the prediction acquisition module is used for acquiring a predicted total comparison value of each comparison combination in the comparison combination set based on a pre-trained comparison prediction model, wherein the comparison prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a comparison combination sample and a total comparison value sample corresponding to the comparison combination sample; and the result determining module is used for determining a plurality of comparison combinations with the largest sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values, and the comparison combinations are used as the matching results of the live-broadcast continuous-cast comparison of the plurality of the hosts to be matched, wherein the same hosts to be matched do not exist in the plurality of comparison combinations.

In a third aspect, an embodiment of the present application provides an electronic device, including: a memory; one or more processors coupled with the memory; one or more applications, wherein the one or more applications are stored in memory and configured to be executed by the one or more processors, the one or more applications configured to perform the live link matching method provided in the first aspect above.

In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, where a program code is stored, where the program code may be called by a processor to perform the live link matching method provided in the first aspect.

According to the live-broadcast continuous-wheat matching method, device, electronic equipment and storage medium, after a plurality of to-be-matched anchors in a matching pool are obtained, a comparison combination set among the plurality of to-be-matched anchors is obtained, the comparison combination set comprises at least one comparison combination, the comparison combination is composed of two different to-be-matched anchors in the plurality of to-be-matched anchors, the predicted total comparison value of each comparison combination in the comparison combination set is obtained based on a pre-trained comparison prediction model, and therefore a plurality of comparison combinations with the largest sum of the predicted total comparison values can be determined from the comparison combination set according to the predicted total comparison value to serve as the matching result of the live-broadcast continuous-wheat comparison of the plurality of to-be-matched anchors, and the same to-be-matched anchor does not exist in the plurality of comparison combinations. According to the application, the total ratio spelling value of various ratio spelling combinations can be predicted more accurately through the pre-trained ratio spelling prediction model, so that a plurality of anchor to be matched in the matching pool can be reasonably matched according to the predicted total ratio spelling value of the various ratio spelling combinations. By combining a plurality of comparison and spelling with the maximum sum of the predicted total comparison and spelling values, as the matching result of the continuous wheat PK of the matching pool, the scheme that the total comparison and spelling values of the matched plurality of comparison and spelling combinations are not too high due to unreasonable pairing can be removed, the problem that the live broadcasting level difference of the paired main players and the main players is large is avoided, meanwhile, the intensity of the live broadcasting continuous wheat PK can be improved, the live broadcasting effect of the continuous wheat PK is improved, the retention rate of each main broadcasting and broadcasting room is further improved, and the conversion rate and the transaction amount of a live broadcasting electric business are also promoted.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the description of the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

Fig. 1 shows an application scenario diagram of a live link matching method provided by an embodiment of the present application.

Fig. 2 shows a flow chart of a live link matching method according to an embodiment of the present application.

Fig. 3 shows a schematic view of an effect provided by an embodiment of the present application.

Fig. 4 is a schematic flow chart of a live link matching method according to another embodiment of the present application.

Fig. 5 shows a flowchart of step S230 in the live link matching method according to the embodiment of the present application.

Fig. 6 shows a flowchart of step S232 in the live link matching method according to the embodiment of the present application.

FIG. 7 illustrates a schematic block diagram of a data flow of a specific-match prediction model suitable for use in embodiments of the present application.

Fig. 8 is a schematic flow chart of step S240 in the live link matching method according to the embodiment of the present application.

Fig. 9 shows a flowchart of a live link matching method according to another embodiment of the present application.

Fig. 10 is a schematic flow chart of step S310 in the live link matching method according to the embodiment of the present application.

Fig. 11 shows a flowchart of step S313 in the live link matching method according to the embodiment of the present application.

FIG. 12 illustrates a schematic block diagram of a data flow of a real-force prediction model suitable for use in embodiments of the present application.

Fig. 13 shows a block diagram of a live-broadcast link matching device according to an embodiment of the present application.

Fig. 14 shows a block diagram of an electronic device according to an embodiment of the present application.

Fig. 15 shows a storage unit for storing or carrying program codes for implementing the live link matching method according to an embodiment of the present application.

Detailed Description

In order to enable those skilled in the art to better understand the present application, the following description will make clear and complete descriptions of the technical solutions according to the embodiments of the present application with reference to the accompanying drawings.

Referring to fig. 1, fig. 1 is a schematic diagram illustrating an application scenario of a live broadcast link matching method according to an embodiment of the present application, where the application scenario includes a live broadcast interaction system 10 according to an embodiment of the present application. The live interaction system 10 includes: terminal device 100 and server 200. Wherein the terminal device 100 and the server 200 are located in a wireless network or a wired network, the terminal device 100 and the server 200 can perform data interaction. In some embodiments, the number of terminal devices 100 may be plural, the server 200 may be communicatively connected to the plurality of terminal devices 100, and the plurality of terminal devices 100 may also be communicatively connected to each other through the internet, or the server 200 may be used as a transmission medium, and implement data interaction between each other through the internet.

In the embodiment of the present application, the terminal device 100 may be a mobile phone, a smart phone, a notebook computer, a desktop computer, a tablet computer, a personal digital assistant (Personal Digital Assistant, PDA), a media player, an intelligent television, a wearable electronic device, etc., and the specific type of the terminal device may not be limited in the embodiment of the present application. The server 200 may be a separate server, may be a server cluster, may be a local server, may be a cloud server, and may not be limited in the embodiment of the present application.

In some embodiments, a client may be installed within the terminal device 100. The client may be a computer Application (APP) installed on the terminal device 100, or may be a Web client, which may refer to an Application developed based on a Web architecture. The user logs in through the account at the client, and all information corresponding to the account can be stored in the storage space of the server 200. The information corresponding to the account comprises information input by a user through the client, information received by the user through the client and the like.

In some implementations, the client may be an application of a live platform, and live content may be displayed on a live interface of the client. The clients may be divided into: a client for anchor use and a client for viewer use. The host may upload the video collected by the local camera to the server 200 through the client used, and then the server 200 forwards the live video to the clients used by all viewers who are in the same channel (or live room) as the host, so that the live video may be displayed on the live interface of the client used by the viewers.

In some embodiments, the live platform may allow multiple hosts to link to perform multi-host interactive live. As a mode, the live broadcast platform can provide a live broadcast interaction mode of cross-channel wheat linking by two or more anchor and comparison (PK event), and the mode can display live broadcast video pictures of channels where the anchor with the wheat is located in a live broadcast interface. In this way, the anchor can play PK events, such as singing, dancing, etc., thereby attracting more viewers.

Further, the client may also receive a trigger event (e.g., a click event, a touch event, etc.) based on the client input by the user, which may act on the manipulation object displayed on the live interface. The client receives the trigger event and can execute the operation corresponding to the control object acted by the trigger event. The control object may be an entire screen displayed on the live interface, or may be display content in the screen. As one way, the manipulation object may be a virtual gift displayed in a screen, and the client used by the viewer may trigger to perform a corresponding operation when receiving a trigger event acting on the virtual gift. For example, interactive data is generated and sent to the server 200, which is forwarded by the server 200 to clients for use by the anchor of the same channel (or live room) and clients for use by other viewers. Therefore, when a user watches live broadcast in the live broadcast room, the virtual gift can be presented to the host of the live broadcast room, and the virtual gift presented by other users can be seen, so that interaction between the audience and the host is realized, and interaction between the audience and the audience is realized. Further, during the live ratio process, the audience can assist (vote) for the corresponding anchor in a manner of presenting the virtual gift, that is, the virtual gift presented by the audience during the live ratio process can be converted into a ratio value (PK value) and added to the ratio bar (PK bar) or integral bar of the presented anchor. If the audience gives a gift of 1 hair money to the anchor, the anchor can increase 1 ticket.

However, in the related art, the live broadcast and link matching mechanism is single, usually random matching, so that the problem of large live broadcast level difference of the two main broadcasters easily occurs, and the live broadcast interaction effect is poor, so that the audience retention rate of a live broadcast room can be improved to a certain extent. Therefore, in order to solve the above-mentioned defects, the embodiment of the application provides a live broadcast continuous wheat matching method, a device, a system, an electronic device and a storage medium. The following will explain in detail specific examples. .

Referring to fig. 2, fig. 2 shows a flow chart of a live link matching method provided by an embodiment of the present application, which may be applied to an electronic device, where the live link matching method may include:

step S110: and obtaining a plurality of anchor to be matched in the matching pool.

In the embodiment of the application, in order to improve the user retention rate of the live broadcasting room, live broadcasting and media Playing (PK) interaction can be performed between the anchor and the anchor. The continuous-play PK interaction can relate to various interactive scenes of specific competition, such as playing games PK, singing songs, dancing, magic, speaking, and the like. When two anchor performs the link-to-microphone PK interaction, the client may display a live interface including the link-to-microphone PK interactions of the two anchor.

For example, referring to fig. 3, fig. 3 shows a live interface schematic diagram in a live ratio, when a live link PK interaction is performed between a live broadcast (i.e., my anchor) of a live broadcast room where a viewer is located and a live broadcast (i.e., hostile anchor) of another live broadcast room, a live video screen 112 of a channel where the my anchor is located, a live video screen 111 of a channel where the hostile anchor is located, and ratio progress information 101 (e.g., link PK time, link PK bar, etc.) may be included in a live broadcast interface 110 displayed by a client used by the viewer. The client side displaying the live interface with the link PK interaction can be a client side used by a spectator or a client side used by a host.

In some embodiments, the live video frames of each link PK host may be laid out separately in the live interface. As a way, the live video pictures of each link-microphone PK anchor can be uniformly laid out. For example, referring to fig. 3, when two hosts perform the link PK interaction, the left half of the live interface displays live video frames 112 of the my host (i.e., the host corresponding to the live broadcasting room where the client is currently located), and the right half displays live video frames 111 of the enemy host (i.e., other link PK hosts except the my host in the link PK interaction). Alternatively, the proportion of the live video pictures of the my anchor in the live interface may be greater than the proportion of the live video pictures of the enemy anchor in the live interface, so that the audience focuses on the live content of the my anchor.

In some embodiments, each live video picture of the live link PK host may occupy the entire live interface, or may occupy only a partial area of the live interface. The specific live interface layout may be set according to the user's requirements, and is not limited herein. For example, referring to fig. 3, the live video screen 112 of the channel where the my anchor is located and the live video screen 111 of the channel where the enemy anchor is located occupy a partial area of the live interface 10, and the other areas 113 may be interaction areas for displaying interaction information, such as a message of sending a gift to a viewer, a chat message of the viewer and the viewer, and the like, and may also be used for operating functional controls for sending gifts, speaking or sharing.

Further, in some embodiments, the PK anchor in the live link PK interaction may be determined based on an invitation mechanism or may be determined based on a matching mechanism. Based on the invitation mechanism, the determination may be that the host corresponding to the live broadcasting room where the client is currently located actively invites other hosts to start the link-to-microphone PK interaction, or that the host corresponding to the live broadcasting room where the client is currently located is invited by other hosts to start the link-to-microphone PK interaction. Based on the matching mechanism, the anchor of the PK objects to be matched can be sent into a matching pool, and the system performs random matching on the anchor of the PK objects to be matched in the matching pool, so that each anchor in the matching pool can be matched with one link-to-wheat PK anchor object.

At present, when the link-microphone PK anchor in the live-broadcast link-microphone PK interaction is determined based on a matching mechanism, the situation that the live-broadcast level difference of the two anchor is quite different is easy to occur due to random matching, so that PK is in a side-to-side state, the irritation to audiences caused by live-broadcast ratio spelling is reduced, the participation degree of the audiences and the audience retention rate of the live-broadcast room of the two anchor are influenced to a certain extent. Therefore, in the embodiment of the application, the total ratio matching value of the various ratio matching combinations can be predicted, so that the plurality of anchor to be matched in the matching pool can be reasonably matched according to the predicted total ratio matching value, a scheme that the total ratio matching value of the matched plurality of ratio matching combinations cannot be too high due to unreasonable matching can be eliminated, and the problem of large live broadcasting level difference of the paired anchor parties can be avoided.

Specifically, a plurality of to-be-matched anchor in the matching pool can be acquired first, so that possible specific tenants can be determined according to the to-be-matched anchor. The anchor to be matched is the anchor of the PK object to be matched.

As a way, when a client used by a host initiates live link-microphone PK interaction in a random matching mode, for example, clicking a corresponding "fight mode-random matching" button in a live page, the client may send a random matching instruction to the server, and when the server receives the random matching instruction, the host ID that is sending the random matching instruction may be put into a matching pool. Of course, if there are multiple anchors initiating a live link PK interaction in a random matching mode, the server may place the anchor IDs of the multiple anchors into the matching pool. Therefore, when a plurality of anchor to be matched in the matching pool is required to be matched according to the matching mechanism, a plurality of anchor IDs existing in the matching pool, namely a plurality of anchor to be matched in the matching pool, can be acquired.

Step S120: and acquiring a comparison and spelling combination set among the plurality of to-be-matched anchor, wherein the comparison and spelling combination set comprises at least one comparison and spelling combination, and the comparison and spelling combination is composed of two different to-be-matched anchor in the plurality of to-be-matched anchor.

In the embodiment of the application, when a plurality of anchor to be matched in the matching pool is obtained, a comparison and spelling combination set among the anchor to be matched can be obtained. Wherein the set of ratio combinations may include at least one ratio combination made up of two different ones of the plurality of principals to be matched.

In some embodiments, any two different principals to be matched in the multiple principals to be matched may be combined to obtain all the pairwise comparison combinations that may exist, that is, a comparison combination set between the multiple principals to be matched is obtained. As a mode, a possible combination scheme of each anchor to be matched and other anchors in the matching pool can be obtained first, and then the duplication elimination processing is carried out, so that a plurality of mutually different comparison and assembly combinations, namely a comparison and assembly set among a plurality of anchors to be matched, can be obtained.

It will be appreciated that a set of comparison combinations between multiple anchor to be matched may be obtained from the definition of the combination. That is, when there are n anchor to be matched in the matching pool, from the n anchor to be matched, 2 anchors are selected and grouped, namely, a comparison combination of 2 anchors selected from the n anchor to be matched for pairing; all the comparison and spelling combinations taken out from the n main broadcasting to be matched are the comparison and spelling combination sets among the plurality of main broadcasting to be matched, and the combination number of the comparison and spelling combination contained in the comparison and spelling combination sets can be n (n-1)/2.

For example, if there is a host A, B, C, D in the matching pool, then there may be a combination scheme (i.e., a set of specific combinations) of: six (A, B), (A, C), (A, D), (B, C), (B, D), (C, D). I.e. the number of combinations is 4*3/2=6.

Step S130: and obtaining a predicted total comparison value of each comparison combination in the comparison combination set based on a pre-trained comparison prediction model, wherein the comparison prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a comparison combination sample and a total comparison value sample corresponding to the comparison combination sample.

When the live broadcasting level difference of the two main broadcasting is too large, PK situation is easy to fall down, so that the stimulation to the audience caused by live broadcasting ratio spelling can be reduced to a certain extent, the enthusiasm of the audience for assisting (voting) the corresponding main broadcasting is low, and further, when the live broadcasting is finished, the total ratio spelling value of the PK and the PK is also low, and the live broadcasting effect of the link PK is not good. Therefore, in the embodiment of the application, after the specific spelling combination set among the plurality of anchor to be matched is obtained, the predicted total specific spelling value of each specific spelling combination in the specific spelling combination set can be obtained based on a pre-trained specific spelling prediction model. And reasonably pairing a plurality of anchor to be matched in the matching pool according to the predicted total ratio spelling value, so that a pairing scheme that the total ratio spelling value of the matched plurality of ratio spelling combinations is not too high due to unreasonable pairing can be eliminated, and the problem of large live broadcasting level difference of the paired anchor and the anchor is avoided. The total ratio value may be the sum of the ratio values presented by two anchor in the ratio combination when the live ratio is over.

In some embodiments, the pre-trained comparison prediction model may be obtained by training the neural network model in advance according to a large number of training samples. Wherein, the training samples can comprise a combination sample and a total combination value sample corresponding to the combination sample. The total ratio spelling value of each ratio spelling combination in the ratio spelling combination set can be predicted according to the pre-trained ratio spelling prediction model, namely the pre-trained ratio spelling prediction model can be used for outputting the predicted total ratio spelling value corresponding to the ratio spelling combination according to the input ratio spelling combination.

In some embodiments, the live platform may record the actual live link PK scene. Therefore, the recorded two parties participating in the live broadcast link microphone PK and the total comparison value presented by the final comparison result can be used as training samples for training the comparison prediction model. And in the history comparison and spelling record, the anchor of the PK in the xth field company wheat PK is taken as a comparison and spelling combination sample, and the total comparison and spelling value presented by the final comparison and spelling result corresponding to the xth field company wheat PK is taken as a total comparison and spelling value sample corresponding to the comparison and spelling combination sample. Wherein x is any number of fields recorded in the history ratio spelling record. Therefore, a training sample set containing a large number of training samples can be obtained according to each continuous wheat PK of the historical record, and further training can be carried out according to the training sample set for comparing and spelling the prediction model, so that the training effect of comparing and spelling the prediction model is ensured.

In some embodiments, the training samples may also be updated based on live link PK recorded in real-time. Therefore, the comparison prediction model can be trained according to the training sample attached to the current actual scene, so that the total comparison value predicted by the trained comparison prediction model accords with the recent trend, and the prediction accuracy of the comparison prediction model is ensured.

Step S140: and determining a plurality of comparison combinations with the maximum sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values, wherein the comparison combinations are used as the matching results of the live-broadcast continuous-wheat comparison of the plurality of the hosts to be matched, and the same hosts to be matched do not exist in the comparison combinations.

In the embodiment of the application, when the predicted total ratio match value of each ratio match in the ratio match set is obtained, a plurality of ratio match combinations with the largest sum of the predicted total ratio match values can be determined from the ratio match set according to the predicted total ratio match value of each ratio match set and used as the matching result of the live-broadcast continuous ratio match of a plurality of ratio match hosts in the current matching pool, wherein the determined plurality of ratio match combinations do not have the same ratio match hosts.

In the embodiment of the application, the total ratio spelling value of the ratio spelling combination can indicate the intensity of the ratio spelling combination, the participation enthusiasm of audience and the live effect of the link-microphone PK. That is, when the total specific spelling value of the specific spelling combination is higher, the specific spelling intensity of the specific spelling combination is higher, the participation enthusiasm of the audience is higher, the assisting force of the audience for the anchor is higher, and the live broadcasting effect of the continuous microphone PK is better. Therefore, the matching result of the PK object of each anchor to be matched in the current matching pool can be determined according to the size sorting of the predicted total ratio value of each ratio combination in the ratio combination set, namely, a plurality of ratio combinations in the current matching pool are determined. The determined multiple ratio combinations do not have the same anchor to be matched, namely a single matching principle.

It can be understood that, because the same anchor cannot exist in the pairwise matching multiple matching combinations, when n anchors to be matched exist in the matching pool, although there are multiple pairwise matching results, each pairwise matching result finally contains n/2 matching combinations. For example, suppose there are 20 anchor waiting for pairing in the match pool, and whichever match results for pairing will contain 10 match-making combinations.

Illustratively, if there are four anchors A, B, C, D in the matching pool, there are six combining schemes: (A, B), (A, C), (A, D), (B, C), (B, D), (C, D). Based on a single matching principle, one possible matching result of pairwise matching is: { (A, B), (C, D) }; another pairwise match that may exist is the result of: { (A, C), (B, D) }, yet another pairwise match that may exist is: { (A, D), (B, C) }. It can be seen that although there are multiple pairwise matches, there is no identical anchor between the pairwise matches in each pairwise match, and there are 4/2=2 pairwise matches in each pairwise match.

Since there may be multiple matching results for pairwise pairs determined from the set of comparison sets, there may be unreasonable matching comparison sets in some of the matching results. Therefore, a reasonable one of the plurality of matching results needs to be screened out to ensure the screened matching result.

In the embodiment of the application, a plurality of specific assembly combinations with the largest sum of the predicted total specific assembly values can be determined from the specific assembly combination set according to the predicted total specific assembly value of each specific assembly combination and used as the matching result of the live broadcast continuous-play specific assembly of a plurality of main broadcasting to be matched in the current matching pool. Therefore, the matching result with the largest sum of the predicted total comparison values is selected and used as the matching result of the live broadcast and the live broadcast connection comparison of the plurality of the hosts to be matched in the matching pool, and the live broadcast effect of each host of the plurality of the hosts to be matched in the matching pool is comprehensively considered, so that the optimization of the whole live broadcast effect is achieved.

It will be appreciated that when there is an unreasonably paired ratio match in a matching result, the matching rationality of other ratio matches in the matching result is also greatly affected, and not only is the overall ratio of the unreasonably paired ratio matches not too high, but the overall ratio of other ratio matches is affected, resulting in the sum of the predicted overall ratio of all ratio matches in the matching result being affected. Because the sum of the predicted total spellings of the matching result is affected, the matching result is not the matching result with the largest sum of the predicted total spellings. That is, by screening out the matching result with the largest sum of the predicted total ratio values, as the matching result of the live broadcast even-wheat ratio of the plurality of the main broadcasters to be matched in the current matching pool, the matching scheme that the sum of the total ratio values of the plurality of the ratio combinations is influenced (usually not too high) due to unreasonable matching can be eliminated, thereby avoiding the problem that the difference between the live broadcast levels of the two main broadcasters is larger, and ensuring the optimization of the whole live broadcast effect.

According to the live-broadcast continuous-wheat matching method provided by the embodiment of the application, after the plurality of to-be-matched anchors in the matching pool are obtained, the comparison combination set between the plurality of to-be-matched anchors is obtained, the comparison combination set comprises at least one comparison combination, the comparison combination is composed of two different to-be-matched anchors in the plurality of to-be-matched anchors, so that the predicted total comparison value of each comparison combination in the comparison combination set is obtained based on a pre-trained comparison prediction model, and therefore, a plurality of comparison combinations with the largest sum of the predicted total comparison values can be determined from the comparison combination set according to the predicted total comparison value to serve as the matching result of the live-broadcast continuous-wheat comparison of the plurality of to-be-matched anchors, wherein the plurality of comparison combinations do not have the same to-be-matched anchor. According to the application, the total ratio spelling value of various ratio spelling combinations can be predicted more accurately through the pre-trained ratio spelling prediction model, so that a plurality of anchor to be matched in the matching pool can be reasonably matched according to the total ratio spelling value. By combining a plurality of comparison and spelling with the largest sum of the predicted total comparison and spelling values, the scheme that the total comparison and spelling values of the plurality of comparison and spelling combinations matched due to unreasonable pairing cannot be too high can be eliminated as the matching result of the direct broadcast and matching of the matching pool, the problem that the direct broadcast and matching of two parties of a main broadcasting has larger level difference is avoided, meanwhile, the intensity of direct broadcast and matching PK can be improved, the direct broadcast and matching effect is improved, the retention rate of each main broadcasting and direct broadcasting room is further improved, and the conversion rate and the transaction amount of direct broadcast and electronic commerce are also promoted.

Referring to fig. 4, fig. 4 shows a flow chart of a live link matching method according to another embodiment of the present application, which may be applied to an electronic device, where the live link matching method may include:

step S210: and obtaining a plurality of anchor to be matched in the matching pool.

In the embodiment of the present application, step S210 may refer to the content of the above embodiment, and is not described herein.

In some embodiments, the number of anchor to be matched in the matching pool may be limited, so as to ensure that the duration of obtaining the matching result is not too long, that is, ensure that the duration of waiting for the anchor and the audience is not too long, and improve the user experience. The number of the anchor to be matched limited in the matching pool can be set reasonably according to the requirement, and is not limited herein. For example, to have the length of time for a match between the anchor and the audience waiting on the order of 1-3 seconds, the number of anchors to be matched in the match pool may be limited to about 20-50.

In some embodiments, when the number of the anchor to be matched that initiates the PK matching request at the same time is relatively large, a plurality of matching pools may be generated, and the number of the anchor to be matched in each matching pool does not exceed the number of the restrictions, so that each matching pool may perform pairwise matching. That is, when the number of the anchor to be matched in any matching pool reaches the limited number, the matching result corresponding to the matching pool can be directly obtained according to the live-broadcast continuous-wheat matching method.

As another implementation mode, when the number of the anchor to be matched in a certain matching pool does not reach the limited number, the corresponding matching result can be obtained directly according to the live-broadcast continuous-wheat matching method after waiting for the specified duration, so that long-time waiting of the anchor and the audience is avoided. That is, after waiting for a specified period of time, if the number of the anchor to be matched in a certain matching pool still does not reach the limited number, the waiting can be stopped, and the corresponding matching result can be obtained directly according to the live-broadcast continuous-wheat matching method. The waiting time period may be set reasonably according to the requirement, and is not limited herein. For example, a wait of 1 minute may be set.

Step S220: and acquiring a comparison and spelling combination set among the plurality of to-be-matched anchor, wherein the comparison and spelling combination set comprises at least one comparison and spelling combination, and the comparison and spelling combination is composed of two different to-be-matched anchor in the plurality of to-be-matched anchor.

Step S230: and obtaining a predicted total ratio matching value of each ratio matching combination in the ratio matching combination set based on a pre-trained ratio matching prediction model, wherein the ratio matching prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a ratio matching combination sample and a total ratio matching value sample corresponding to the ratio matching combination sample.

In the embodiment of the present application, the steps S210 to S230 may refer to the content of the above embodiment, and are not described herein.

In some embodiments, historical spell features may be extracted from the anchor's historical spell records, and training may be performed against the spell prediction model. Therefore, useful information of the comparison prediction model can be extracted from the historical comparison record, the original data is converted into the characteristics, the actual problem of the comparison prediction model processing can be better represented, the accuracy of unknown data is improved, and the training effect of the comparison prediction model is guaranteed.

Specifically, referring to fig. 5, step S230 may include:

step S231: the method comprises the steps of obtaining a first historical comparison feature of a first anchor in a current comparison combination and a second historical comparison feature of a second anchor in the current comparison combination, wherein the current comparison combination is any comparison combination in the comparison combination set.

In some embodiments, after a set of ratio combinations is obtained, a prediction of the total ratio value may be made for any one of the ratio combinations. Specifically, a first historical spell feature of a first anchor in a current spell combination and a second historical spell feature of a second anchor in the current spell combination may be obtained. The current ratio combination is any ratio combination in the ratio combination set, and the first anchor and the second anchor are two different anchors to be matched paired in the current ratio combination. Thus, the total ratio value prediction of each ratio combination of the ratio combination set is traversed, and the predicted total ratio value of each ratio combination is obtained.

In some embodiments, the history matching features may include a consumption capability feature of a user of the anchor live room, an anchor history matching value feature, and an assistance feature of the user of the anchor live room during the anchor history matching. The consumption capability feature of the user can be used for representing payment conditions of the user on the live client, the anchor history comparison value feature can be used for representing average comparison value conditions of recent live link and comparison field times of the anchor, and the boosting feature of the user in the anchor history link and comparison process of the anchor living room can be used for representing boosting conditions (namely gift giving conditions or voting conditions) of all users in the living room on the anchor recently.

Because users may pay for various reasons on a live client, and the consumption of different users may vary, the consumption may be a 0 to millions gap, in some embodiments, the user's payment may be ranked to facilitate training with an input model, with different ranks representing different consumption capabilities of the user. For example, the payment may be divided into five categories A, B, C, D, E, with the amount of the consumption decreasing stepwise from A to E. If the grade E represents consumption of 0-1 yuan, the grade A represents consumption of more than 100000 yuan, etc. Therefore, the consumption capability characteristics of the users in the live broadcasting room can be obtained according to the payment level conditions of the users in the live broadcasting room.

As a way, the distribution of the users of the live broadcasting room in each payment level can be counted and converted into the consumption capability characteristics of the users of the live broadcasting room. For example, there are 100 users in the current hosting live broadcast room, and the distribution situation of the users at each payment level is: a-1, B-10, C-20, D-40 and E-29. Whereby the consumer capacity of the users of the hosting direct broadcast room is characterized by [1,10,20,40,29].

In some embodiments, the average ratio value of the recent live ratio of the anchor may be determined based on the final ratio value of each live link of the anchor within a preset time period before the current. As one way, the average ratio of recent live ratio shots of the anchor may be the ratio of the sum of the final ratio of each shot to the total shots within a preset time period, e.g., the sum of recent 3 shot PK values: 14000+12000+10000=36000, average ratio value=36000/3=12000. As another way, corresponding weights can be set for each live broadcast ratio spelling in the preset duration, the closer the live broadcast ratio spelling is to the current time, the higher the weights are, so that the history ratio spelling characteristics of the anchor are ensured to be more in line with the recent actual conditions, and the accuracy of model prediction is improved. The preset time length can be set reasonably according to requirements, and is not limited in the embodiment of the application. For example, one month, half month, one week, etc.

In some embodiments, the boosting feature of the users of the hosting direct broadcast room in the hosting history matching process can be understood as a desired payment condition of all the users of the hosting direct broadcast continuous matching process. In one mode, the total payment amount of the target user in the live broadcast matching process in the current previous specified time period can be obtained, the total scene of the target user watching the live broadcast matching of the live broadcast and the live broadcast matching of the live broadcast can be obtained, and the ratio of the total payment amount to the total scene is used as the expected payment of the target user for the live broadcast matching of the live broadcast of each live broadcast. The target user is any user in the live broadcasting room. And then, according to the expected payment of all users, acquiring the sum of the expected payment of all users for the live broadcast continuous-wheat comparison of the host broadcast as the power assisting characteristic of the users in the host broadcast historic continuous-wheat comparison process of the host broadcast living broadcast room. The specified duration may be set reasonably according to requirements, which is not limited in the embodiment of the present application. For example, one month, half month, one week, etc. As a way, the specified duration may be consistent with the preset duration to ensure that there is a correlation between the user assistance feature and the anchor history comparison value feature.

For example, user a views 10 live link to the host b, where 5 virtual gifts (pay-through Fei Chuli) are presented to the host b, and the total amount of the 5 virtual gifts presented is W. User a expects to pay =w/10 for each live link PK of the host.

The method comprises the steps of obtaining a first anchor and a second anchor paired in a current ratio matching combination, then respectively obtaining the consumption capability characteristic, the first anchor history ratio matching value characteristic and the assistance characteristic of the user in the first anchor history ratio matching process of the first anchor living broadcast room, splicing to obtain the first history ratio matching characteristic of the first anchor, obtaining the consumption capability characteristic, the second anchor history ratio matching value characteristic and the assistance characteristic of the user in the second anchor living broadcast room in the second anchor history ratio matching process, and splicing to obtain the second history ratio matching characteristic of the second anchor. Repeating the above operation, so as to traverse each specific assembly of the specific assembly set, and obtain a first historical specific assembly characteristic of a first anchor in each specific assembly and a second historical specific assembly characteristic of a second anchor in each specific assembly. The history comparison feature can be a feature matrix formed by [ X, Y, Z ], wherein X refers to the consumption capability feature of the user in the anchor living broadcast room, Y refers to the anchor history comparison value feature, and Z refers to the assistance feature of the user in the anchor living broadcast room in the anchor history comparison process.

Step S232: inputting the first historical comparison feature and the second historical comparison feature into a pre-trained comparison prediction model to obtain a predicted total comparison value corresponding to the current comparison combination output by the comparison prediction model, wherein the comparison prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a historical comparison feature sample corresponding to the comparison combination sample and a total comparison value sample corresponding to the historical comparison feature sample.

In some embodiments, after the first historical specific feature of the first anchor in the current specific combination and the second historical specific feature of the second anchor in the current specific combination are obtained, the first historical specific feature and the second historical specific feature can be input into a specific prediction model trained in advance, so as to obtain a predicted total specific value corresponding to the current specific combination output by the specific prediction model, namely, the sum of the predicted specific values of the first anchor and the second anchor.

In some embodiments, the above-mentioned piecewise prediction model may be a deep neural network (DNN, deep Netural Network) model, so that higher-order information and cross information of features can be explored, feature combinations that do not occur in training samples can be better generalized, and accuracy of prediction is improved. Wherein the piecewise predictive model may include multiple hidden layers. The meaning of the hidden layer is to abstract the input features to another dimension space to display the more abstract features, and the features can be better linearly divided. The multi-layer hidden layer is used for carrying out multi-layer abstraction on the input characteristics to obtain the high-order information of the characteristics, so that different types of data can be divided better in a linear manner.

As one approach, the comparative prediction model may include an input layer, a 3-layer hidden layer, and an output layer, where the 3-layer hidden layer dimensions are 128, 32, 4, respectively. Therefore, the first history specific spelling characteristic and the second history specific spelling characteristic can be input into the 3-layer hidden layer through the input layer to carry out deep learning prediction, and the output layer outputs the predicted total specific spelling value corresponding to the current specific spelling combination.

In some embodiments, since different data may have different data dimensions, to facilitate prediction of the subsequent neural network model, an embedding layer may be added before the hidden layer, i.e., before the first and second history spell features are input into the hidden layer, the first and second history spell features are input into the embedding layer for feature embedding (Feature Embedding) to transform (or reduce dimensions) the first and second history spell features into a fixed-size feature representation (vector).

In some embodiments, because the consumption capability feature of the user of the anchor live broadcast room in the history comparison feature, the anchor history comparison value feature, and the assistance feature of the user of the anchor live broadcast room in the anchor history comparison process are all continuous features, before the history comparison feature is input into the embedded layer, the history comparison feature may be subjected to a barrelling process to convert the continuous features into discrete features.

In some embodiments, the pre-trained predictive model may be trained from a large number of training samples. The training samples can comprise historical comparison characteristic samples corresponding to the comparison combination samples and total comparison value samples corresponding to the historical comparison characteristic samples. The pre-trained comparison and spelling prediction model can be used for outputting a prediction total comparison and spelling value corresponding to the first history comparison and spelling characteristic and the second history comparison and spelling characteristic according to the obtained first history comparison and spelling characteristic and the second history comparison and spelling characteristic corresponding to the current comparison and spelling combination.

The training samples can be obtained through the appointed processing of the history comparison records of each anchor. Specifically, the history comparison records of each anchor each field are collected in advance, the history comparison records of each field comprise consumption capability features of users in an anchor live room before comparison, anchor history comparison value features before comparison, power assisting features of users in an anchor history comparison process of the anchor live room before comparison and total comparison values of comparison parties at the end of comparison, and the history comparison feature corresponding to each field history comparison record can be obtained by splicing the consumption capability features of the users in each field history comparison record, the anchor history comparison value features and the power assisting features of the users in the anchor history comparison process, so that the history comparison feature corresponding to each field history comparison record can be used as a history comparison feature sample, the total comparison value corresponding to the field history comparison record is used as a total comparison value sample corresponding to the history comparison feature sample, and the history comparison feature sample and the total comparison value sample are used as a group of training samples and added into a training sample set. So that multiple groups of training samples can be extracted from the training sample set for training by comparing the spelling prediction model one by one.

In some embodiments, historical spell feature samples in a set of training samples may be used as inputs to a spell prediction model, and the total spell value samples in the set of training samples may be used as desired outputs (i.e., learning targets) of the spell prediction model, such that the spell prediction model may be trained once by the actual and desired outputs of the spell prediction model. That is, the historical specific feature samples in the set of training samples can be input into the specific prediction model for forward calculation, so as to obtain the actual output of the specific prediction model, wherein the actual output is the total specific value predicted by the specific prediction model. Because the total comparison value sample in the training samples is used as the expected output of the comparison prediction model, the model parameters can be updated according to the error of the predicted total comparison value and the total comparison value sample, and the pre-trained comparison prediction model is obtained through a large number of iterative training.

In an actual application scene, because the active time of the audience on the live broadcast platform is usually indefinite or regular, the retention rate of users in the live broadcast room of the host is also high and low, that is, the live broadcast effect of the host under different live broadcast time has a larger gap, so that the continuous cropping practice force of the host under different live broadcast time also has a larger gap, therefore, in some embodiments, in order to ensure the prediction accuracy of the continuous cropping prediction model, features such as host identification, continuous cropping PK time and the like can be introduced into the continuous cropping prediction model for prediction.

Specifically, referring to fig. 6, step S232 may include:

step S2321: and acquiring a first anchor identifier of the first anchor and a second anchor identifier of the second anchor.

Step S2322: and obtaining the matching start time of the live broadcast and the wheat connecting comparison.

In some embodiments, a first anchor identifier of a first anchor in a current anchor combination, a second anchor identifier of a second anchor in the current anchor combination, and a matching start time of the live-broadcast link anchor can be obtained, so as to accurately predict anchor forces of the first anchor and the second anchor under the matching time.

Wherein, the anchor identifier can uniquely correspond to one anchor, and different anchor users can be distinguished according to the anchor identifier. In some embodiments, the live identification may be a client identification, a live room number, a host ID account logged on to the client, a host nickname, etc., without limitation.

In some embodiments, the matching start time of the live link matching may be a time point when a client used by the anchor initiates a matching request, or may be a time point when a matching pool is generated, which is not limited herein. As an implementation mode, the matching start time of the live broadcast continuous-time-to-wheat ratio spelling can be obtained in a fixed mode. Wherein, the fixed form can be a form of T hours and M minutes, so that the hour characteristic and the minute characteristic can be extracted respectively. For example, when the matching start time is 12:30, the hour characteristic is 12, and the minute characteristic is 30.

Step S2323: and performing cross feature operation on the first anchor identifier, the second anchor identifier, the matching start time, the first history comparison feature and the second history comparison feature to obtain a first combination feature.

In some embodiments, the obtained first anchor identifier, the second anchor identifier, the matching start time, the first history specific spelling characteristic and the second history specific spelling characteristic may be input into a cross (cross conversion) layer to perform cross characteristic operation, so as to obtain a first combination characteristic. Wherein the cross layer can be used to automatically construct limited high-order cross features. Thus, the multi-dimensional characteristic combining the history comparison characteristic of the first anchor and the second anchor and the multi-dimensional characteristic of the matching starting time are obtained, and the relevance characteristic between the first anchor and the second anchor is obtained. The interaction condition between the historical comparison and spelling characteristics of the first anchor and the second anchor at the matching starting time can be obtained by performing the cross characteristic operation, so that the influence of the cross characteristic on the prediction total comparison and spelling value can be obtained, namely the non-linearity is increased for the comparison and spelling prediction model, and the prediction effect of the model is improved.

In some embodiments, the performing the cross-feature operation may be performing the cross-feature operation on the same type of feature in the first anchor and the second anchor. Specifically, the first anchor identifier and the second anchor identifier may be subjected to cross feature operation, the first history comparison feature and the second history comparison feature are subjected to cross feature operation, the hour feature and the minute feature in the matching start time are subjected to cross operation, and then the cross feature operations obtained respectively are spliced and combined to obtain a first combined feature, so that the first combined feature is sent to the comparison prediction model for prediction. In other embodiments, the hour feature and the minute feature in the matching start time are not subjected to the cross operation, and the matching start time can be directly sent to the piecewise prediction model for prediction.

In some embodiments, the two features perform cross feature operation, which may be that a hash operation is performed on a cartesian product of the two features, and then a hash result is modulo a model set hash_bucket_size, where the obtained result is the cross feature. Wherein the Cartesian product of two features is a feature combination formed by multiplying two feature matrices. The hash operation can compress the high-dimensional feature vector obtained by Cartesian product into a lower-dimensional feature vector, and the expression capacity of the original feature is not lost as much as possible. The hash_bucket_size modular sampling is used for guaranteeing that the characteristic length of the hash result after modular sampling is fixed, and the characteristic dimension can be reduced. Where hash_bucket_size is a modulo length, such as 100.

In some embodiments, since different data may have different data dimensions, to facilitate the cross feature operation, the first anchor identifier, the second anchor identifier, the matching start time, the first history spell feature, and the second history spell feature may be input to the embedding layer to perform feature embedding, so as to convert (or reduce the dimension of) all features into a feature representation (vector) with a fixed size.

Step S2324: and inputting the first combined characteristic into a pre-trained comparison prediction model.

In some embodiments, after the first combined feature is obtained, the first combined feature may be input into a pre-trained ratio-matching prediction model to predict, so as to obtain a predicted total ratio value of the current combined ratio output by the ratio-matching prediction model.

In some embodiments, since the dimensions of the combined features after the cross feature operation are too high, the features after the cross feature operation need to be input to the embedding layer again for feature embedding to perform data dimension reduction (for example, dimension reduction to 10). I.e. an embedded layer is added between the cross layer and the hidden layer.

The present embodiment is exemplarily described below with reference to fig. 7:

referring to fig. 7, fig. 7 is a schematic diagram illustrating a data flow of a comparison prediction model according to an embodiment of the present application. As shown in fig. 7, the features of the live broadcasting room of the first anchor and the second anchor in the comparison and spelling combination and the matching start time are input into the embedding layer for feature embedding, then the features of the live broadcasting room of the first anchor and the second anchor after embedding and the matching start time are input into the cross layer for cross combination, then a 3-layer neural network (the feature dimensions of the 3-layer hidden layer are 128, 32 and 4 respectively) is input for deep learning, and finally a prediction of the total comparison and spelling value of the first anchor and the second anchor in the comparison and spelling combination is output.

Step S240: and determining a plurality of comparison combinations with the largest sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values and a cluster search algorithm, and taking the comparison combinations as the matching results of the live-broadcast and continuous-cast comparison of the plurality of the hosts to be matched, wherein the same hosts to be matched do not exist in the comparison combinations.

Because there may be multiple matching results for the pairwise pairing determined from the comparison combination set, a matching result with the largest sum of predicted total comparison values needs to be accurately screened from the multiple matching results. In some embodiments, all pairwise matching results may be obtained first, and then the sum of the predicted total comparison values of all comparison combinations in each matching result may be calculated respectively, so that a matching result with the largest sum of the predicted total comparison values may be selected from the sums of the predicted total comparison values corresponding to each matching result, and the matching result may be used as the matching result of the live-broadcast even comparison of a plurality of hosts to be matched in the current matching pool.

When the number of the anchor to be matched in the matching pool is too large, as the matching results of all pairwise matching are too large, if all the matching results are traversed to find the matching result of the sum of the predicted total comparison values, huge calculation amount is generated, and the matching waiting time of the anchor and the audience is seriously prolonged, so that the user experience is poor. For example, assuming that 20 anchors wait for pairing in the matching pool, there are more than 6 hundred million matching results of 5 millions of pairings, and in reality, the number of anchors waiting for matching at the same time is generally about 50, so that it is not practical to traverse all the matching results to find the matching result with the largest sum of the predicted total comparison values.

Therefore, in other embodiments, a plurality of specific matching combinations with the largest sum of the predicted total specific matching values, that is, a matching result with the largest sum of the predicted total specific matching values, can be approximately determined from the specific combination set according to the predicted total specific matching value of each specific matching combination and a Beam Search (Beam Search) algorithm, and the calculation amount can be greatly reduced as the matching result of the live-broadcast continuous specific matching of a plurality of main broadcasting to be matched.

The Beam Search uses breadth-first strategy to build Search tree, at each layer of tree, the nodes are ordered according to heuristic cost, then only the predetermined number (Beam Width-bundling Width) of nodes is left, only these nodes continue to expand at the next layer, and other nodes are cut off. The bundle width is understood as the number of nodes reserved only for each layer.

Because the anchor with greater than collage is typically also higher than collage, in some embodiments, multiple anchor to be matched in the matching pool may be first ranked for specific collage to search for matches from the specific collage ranking size. Specifically, referring to fig. 8, step S240 may include:

step S241: and acquiring the first i comparison combinations corresponding to the target anchor from the comparison combination set according to the order of the predicted total comparison values from large to small, and taking the first i comparison combinations as current i optimal solution sequences, wherein i is the bundling width, the target anchor has the highest strength reference value in the anchor to be matched, and the strength reference value is used for representing the strength of the anchor to be matched.

In some embodiments, a target anchor with the highest strength reference value in a plurality of anchor to be matched in the matching pool may be obtained, so as to prioritize the first anchor to pair according to the strength reference value. The strength reference value is used for representing the strength of the campaigns to be matched, namely, the value limit of the virtual gift presented by the audience can be obtained by the anchor. It will be appreciated that the higher the campaigns' strength of the host to be matched, the more often the audience in its direct broadcast room will be, and thus the higher the specific spelling value of the host.

In some embodiments, the target anchor may be paired with other anchors in the matching pool one by one to obtain all the specific assembly combinations corresponding to the target anchor, and then all the specific assembly combinations corresponding to the target anchor may be predicted according to the specific assembly prediction model, so as to obtain a predicted total specific assembly value of each specific assembly combination corresponding to the target anchor. And then, according to the bundling width i in the bundling search algorithm and the sequence from the big to the small of the predicted total ratio spelling value, the first i ratio spelling combinations which are corresponding to the target anchor and are ranked at the front are obtained from all ratio spelling combinations corresponding to the target anchor and are used as the i maximum optimal solution sequences reserved in the hierarchy. At this time, each optimal solution sequence has only one specific combination, i.e. i specific combinations and i optimal solution sequences are in one-to-one correspondence. The bundling width i indicates that only the largest i comparison and spelling combinations are reserved each time, and the i comparison and spelling combinations corresponding to the target anchor can be understood as i priority comparison and spelling objects of the target anchor.

In some embodiments, each of the set of specific combinations may be predicted in advance according to the above-described specific prediction model to obtain a predicted total specific value for each of the set of specific combinations. Then, all the specific assembly combinations containing the target anchor can be obtained from the specific assembly combination set, and the predicted total specific assembly value of each specific assembly combination is obtained in advance, so that the first i specific assembly combinations which are corresponding to the target anchor and are ranked at the front can be obtained from all the specific assembly combinations containing the target anchor according to the sequence from the large to the small of the predicted total specific assembly value, and serve as the i optimal solution sequences which are reserved the largest in the hierarchy.

For example, when the bundle width i is 3, the anchor 1 is preferentially ordered according to the strength reference value to pair, so that the first 3 ratio-assembled combinations with the top prediction total ratio-assembled value ordering corresponding to the anchor 1 can be (1, 2, 10000), (1, 4, 9500), (1, 5, 9000), which are the largest 3 optimal solution sequences reserved as the current level. Wherein, (1, 2, 10000) has the following meanings: (actual reference value ranks the anchor 1, ranks the anchor 2, and the predicted total value of the combination of the two anchors). It can be seen that the priority carousel object of rank 1 is the carousel of ranks 2, 4, 5.

Step S242: and obtaining the first i ratio combining combinations corresponding to the appointed anchor with the highest strength reference value in the residual ratio combining set according to the sequence from the big to the small of the predicted total ratio combining value, wherein the residual ratio combining set is the ratio combining combination of the anchor to be matched in the target sequence, the residual anchor to be matched is the residual anchor except the anchor to be matched in the target sequence, and the target sequence is any sequence in the i optimal solution sequences.

In some embodiments, after the i best solution sequences that were retained the largest before are obtained, a next level of expansion may be performed for each best solution sequence. Specifically, one sequence can be arbitrarily selected from the i optimal solution sequences as a target sequence, so as to perform next-level expansion of the target sequence, namely, pairing of other anchor except the target sequence. Since the same anchor (single matching principle) cannot appear in the final matching result, the rest anchors to be matched except for the anchors to be matched in the target sequence are acquired from a plurality of anchors to be matched in the matching pool. And then, selecting the designated anchor with the highest strength reference value from the rest anchor to be matched again for preferential pairing.

The description of the preferred pairing process for the designated anchor is similar to that described above for the target anchor preferred pairing process. That is, in some embodiments, the designated anchor may be paired with other anchors in the remaining anchors to be matched one by one to obtain all the specific anchor corresponding to the designated anchor, and then all the specific anchor corresponding to the designated anchor may be predicted according to the specific anchor prediction model, so as to obtain the predicted total specific value of each specific anchor corresponding to the designated anchor. And then, according to the bundling width i in the bundling search algorithm and the sequence from the big to the small of the predicted total ratio spelling value, the first i ratio spelling combinations which are corresponding to the appointed anchor and are ranked at the front are obtained from all ratio spelling combinations corresponding to the appointed anchor. Thus, i priority comparison objects of the designated anchor are obtained.

That is, in other embodiments, since the same anchor cannot appear in the final matching result, the specific assembly that does not include the anchor to be matched in the target sequence may be obtained from the specific assembly set first, and the remaining specific assembly is obtained. And predicting each of the remaining ratio combining groups in advance according to the ratio combining prediction model so as to obtain a predicted total ratio combining value of each of the remaining ratio combining groups. Then, all the specific spelling combinations including the appointed anchor can be obtained from the rest specific spelling combinations, and the first i specific spelling combinations which are corresponding to the appointed anchor and are ranked in front can be obtained from all the specific spelling combinations including the appointed anchor according to the ranking from big to small of the predicted total specific spelling values because the predicted total specific spelling value of each specific spelling combination is obtained in advance.

For example, assuming that (1, 2, 10000) is selected from the above-mentioned 3 optimal solution sequences (1, 2, 10000), (1, 4, 9500) and (1, 5, 9000) as the target sequence to perform the next-level expansion, since the anchor 1 and 2 of the anchor ranking of the actual reference value are already paired, the designated anchor with the highest actual reference value may be selected again from the remaining anchors other than the anchor 1 and 2, that is, the anchor 3 of the anchor ranking of the actual reference value is preferentially paired. The 3 rd anchor is matched and combined with other anchors in the rest anchors one by one, and the first 3 comparison and spelling combinations with the predicted total comparison and spelling value corresponding to the 3 rd anchor being ranked at the front are reserved, such as (3, 4, 5200), (3, 5, 4500), (3, 7, 4000). It can be seen that, based on matching the rank 1 st and rank 2 nd anchor, the rank 3 rd anchor's preferred anchor object may be the rank 4 th, rank 5 th and rank 7 anchor.

It can be understood that, since the current hierarchy retains the largest i optimal solution sequences, the expansion operation of the next hierarchy can be performed for each of the i optimal solution sequences. I.e. the first i specific spelling combinations corresponding to the appointed anchor in each optimal solution sequence are obtained.

For example, for (1, 4, 9500) out of the 3 best solution sequences exemplified above, the first 3 specific combinations of the designated anchor with the highest strength reference among the remaining anchors, i.e., anchor rank 2, may be (2, 3, 6100), (2, 5, 5500), (2, 6, 5000). It can be seen that, based on matching the ranks 1 st and 4 th anchors, the preferred anchor object of the rank 2 nd anchor may be the ranks 3, 5, and 6 anchor.

For another example, for (1, 5, 9000) of the 3 best solution sequences exemplified above, the first 3 kinds of the specific anchor with the highest strength reference among the remaining anchors, i.e., anchor rank 2, may be (2, 3, 6100), (2, 4, 5600), (2, 6, 5000). It can be seen that, based on matching the rank 1 st and rank 5 th anchors, the preferred anchor object of rank 2 nd anchor may be rank 3 rd, 4 th and rank 6 anchor.

Step S243: and splicing the first i comparison and splicing combinations corresponding to the designated anchor with the target sequence respectively to obtain i combination sequences corresponding to the target sequence.

In some embodiments, when the target sequence is selected to perform expansion of the next level, after the first i comparison and spelling combinations corresponding to the designated anchor are obtained, the first i comparison and spelling combinations corresponding to the designated anchor are respectively spliced with the target sequence, so as to obtain i combination sequences corresponding to the target sequence, that is, i expansion of the next level of the target sequence.

For example, when the target sequence is (1, 2, 10000) in the above-exemplified 3 optimal solution sequences, since the anchor is designated as the anchor of rank 3 and the first 3 ratio-splice combinations of the predicted total ratio-splice value rank corresponding to the anchor are (3, 4, 5200), (3, 5, 4500), (3, 7, 4000), the first 3 ratio-splice combinations corresponding to the anchor can be spliced with the target sequence (1, 2, 10000), respectively, and the 3 combined sequences corresponding to the obtained target sequence are: (1, 2, 10000) - (3, 4, 5200); (1, 2, 10000) - (3, 5, 4500); (1,2, 10000) - (3,7, 4000). I.e., (1, 2, 10000) of 3 extension pairs.

For another example, when the target sequence is (1, 4, 9500) in the 3 best solution sequences exemplified above, since the anchor is designated as the anchor of rank 2 and the first 3 specific combinations of the predicted total specific combinations corresponding to the anchor are designated as (2, 3, 6100), (2, 4, 5600) and (2, 6, 5000), the first 3 specific combinations corresponding to the anchor can be spliced with the target sequence (1, 4, 9500), respectively, and the 3 combined sequences corresponding to the obtained target sequence are: (1, 4, 9500) - (2, 3, 6100); (1, 4, 9500) - (2, 4, 5600); (1,4, 9500) - (2,6, 5000). I.e., (1, 4, 9500).

Step S244: and according to the sum of the predicted total comparison values of the i combined sequences corresponding to each sequence in the i optimal solution sequences, acquiring the first i combined sequences as new i optimal solution sequences according to the sequence from large to small.

Since each optimal solution sequence expands i combined sequences, i combined sequences can be obtained by the i optimal solution sequences, and if the next-level expansion is performed, i combined sequences … … can be obtained, so that the combined sequences are gradually increased, and the calculated amount is increased. Therefore, in some embodiments, the i combined sequences corresponding to each of the i optimal solution sequences, that is, i×i combined sequences, may be first obtained, and then the sum of the predicted total ratio values of all ratio combinations in the i×i combined sequences is calculated, and the first i combined sequences with the top order are obtained as new i optimal solution sequences according to the order of the sum of the predicted total ratio values from the top to the bottom. Therefore, each screening is guaranteed to be carried out based on the principle that the predicted total comparison value is forward, and further the fact that the sum of the predicted total comparison values of the finally obtained matching results is also approximately forward is guaranteed, and accordingly the live broadcast effect of each anchor is comprehensively considered, and the overall live broadcast effect optimization is achieved.

For example, in the above-exemplified 3 optimal solution sequences (1, 2, 10000), (1, 4, 9500), and (1, 5, 9000), after each sequence is extended by 3 combined sequences, a total of 3*3 =9 combined sequences are obtained:

first kind: (1, 2, 10000) - (3, 4, 5200), the sum of predicted total spellings: 10000+5000=15200;

second kind: (1, 2, 10000) - (3, 5, 4500), the sum of predicted total spell values: 10000+4500=14500;

third kind: (1, 2, 10000) - (3, 7, 4000), the sum of predicted total spellings: 10000+4000=14000;

fourth kind: (1, 4, 9500) - (2, 3, 6100), predicting the sum of total spell values: 9500+6000=15600;

fifth: (1, 4, 9500) - (2, 5, 5500), the sum of predicted total spellings: 9500+5500=15000;

sixth: (1, 4, 9500) - (2, 6, 5000), predicting the sum of total spell values: 9500+5000=14500;

seventh: (1, 5, 9000) - (2, 3, 6100), predicting the sum of total spellings: 9000+6000=15100;

eighth: (1, 5, 9000) - (2, 4, 5600), predicting the sum of total spell values: 9000+5600=14600;

ninth, the method comprises: (1, 5, 9000) - (2, 6, 5000), predicting the sum of total spell values: 9000+5000=14000;

according to the sequence from the big to the small of the sum of the predicted total ratio values, the first 3 combined sequences with the largest sum of the predicted total ratio values are obtained: (1, 4, 9500) - (2, 3, 6100); (1, 2, 10000) - (3, 4, 5200); (1, 5, 9000) - (2, 3, 6100) as new i optimal solution sequences.

Step S245: and repeating the steps from the rest ratio combining to the first i ratio combining corresponding to the appointed anchor with the highest strength reference value in the rest to-be-matched anchor according to the sequence from the big to the small of the predicted total ratio combining value, and obtaining the first i combined sequences as new i optimal solution sequences according to the sum of the predicted total ratio combining values of the i combined sequences corresponding to each sequence in the i optimal solution sequences, until the i optimal solution sequences contain the plurality of anchor to be matched or only do not contain one anchor in the plurality of anchor to be matched, thereby obtaining final i optimal solution sequences.

In some embodiments, after the new i optimal solution sequences are obtained, the process from step S242 to step S245 may be repeated, that is, for each of the new i optimal solution sequences, the next-level expansion is continued until the i optimal solution sequences include multiple anchor to be matched in the matching pool or only one anchor not including multiple anchor to be matched (iterated n/2 times in total), so as to obtain i optimal solution sequences that are finally expanded by multiple levels.

It can be understood that when the i optimal solution sequences include a plurality of anchors to be matched in the matching pool, all anchors to be matched in the matching pool can be considered to be matched, and when the i optimal solution sequences do not include only one anchor of the plurality of anchors to be matched, the anchors to be matched in the matching pool are described as odd, and one anchor cannot be matched. In some embodiments, when only one anchor remains unmatched in the matching pool, the anchor may be forwarded to other matching pools for matching.

Step S246: and obtaining a sequence with the maximum sum of the predicted total comparison values from the final i optimal solution sequences, and taking the sequence as a matching result of the live-broadcast continuous-wheat comparison of the plurality of main broadcasting to be matched.

In some embodiments, when the final i optimal solution sequences are obtained, a sequence with the largest sum of the predicted total ratio values may be obtained from the final i optimal solution sequences, and the sequence may be used as a matching result of the live-broadcast even-wheat ratio of the plurality of the hosts to be matched. Thereby ensuring approximate optimization of the overall live effect.

Because the expansion of each layer is ensured to keep the sequence with the front sum of the predicted total comparison values by the cluster search algorithm, the obtained i optimal solution sequences are the front sum of the predicted total comparison values of the matching results of the obtained i pairwise pairings, and the sequence with the not too high sum of the predicted total comparison values caused by unreasonable pairing can be eliminated, so that the problem of larger live broadcasting level difference of paired broadcasters and broadcasters is avoided. In addition, the sequence with the largest sum of the predicted total comparison values is obtained from the final i optimal solution sequences with the front sum of the predicted total comparison values and is used as the matching result of the live-broadcast continuous-wheat comparison of the plurality of the hosts to be matched, so that the matching result of the live-broadcast continuous-wheat comparison in the current matching pool can be ensured to be the optimal result of the cluster search algorithm, and the optimization is approximately achieved. It can be understood that when the matching result of the live-broadcast continuous-match matching in the current matching pool is the matching result with the largest sum of the predicted total matching values in all the matching results of the possible pairwise matching, the optimization in the ideal situation can be considered.

In some embodiments, the cluster width i can be enlarged, so that the sequence with the largest sum of the predicted total ratio values in the i finally obtained optimal solution sequences is more close to the optimal sequence, and the overall live broadcast effect is improved. And the time complexity does not exceed O (ni ² ) Where n is the number of anchor and i is the bundle width.

In some embodiments, the strength reference of the anchor may also be predicted by a neural network model. Specifically, referring to fig. 9, before step S241, the live link matching method of the present application may further include:

step S310: and obtaining a predicted strength reference value of each anchor to be matched in the plurality of anchors to be matched based on a pre-trained strength prediction model, wherein the strength prediction model is obtained by training a neural network model according to an anchor sample and a strength reference value sample corresponding to the anchor sample.

In some embodiments, a predicted strength reference value for each of a plurality of to-be-matched anchor in a matching pool may be obtained based on a pre-trained strength prediction model. The strength prediction model can be obtained by training the neural network model according to the anchor sample and the strength reference value sample corresponding to the anchor sample.

In some embodiments, the pre-trained strength prediction model may be obtained by training the neural network model according to a large number of training samples in advance. The training samples may include a anchor sample and a strength reference value sample corresponding to the anchor sample. So as to predict the actual force reference value of each anchor to be matched in the matching pool according to a pre-trained actual force prediction model, that is, the pre-trained strength prediction model can be used for outputting a predicted strength reference value corresponding to the anchor according to the input anchor characteristics.

The ratio value of the host in the live ratio is calculated according to the audience boosting degree, namely the value limit of the virtual gift, and the value limit of the virtual gift presented by the audience just reflects the revenue capacity of the host, so the ratio value of the host in the live ratio is the representation of the strength reference value of the host. Thus, in some embodiments, the resulting ratio value in the anchor history live ratio may be used as a strength reference value sample. Specifically, the live broadcast platform can record the actual live broadcast ratio of the host broadcast. Therefore, the recorded live broadcast and live link comparison and spelling anchor and the comparison and spelling value presented by the final comparison and spelling result of the anchor can be used as a training sample for training the strength prediction model. And in the history comparison and spelling record, any one of the anchor of PK in the x-th live broadcast comparison and spelling is taken as an anchor sample, and the comparison and spelling value presented by the anchor end comparison and spelling result is taken as a strength reference value sample corresponding to the anchor sample. Of course, the other anchor of the PK in the x-th live ratio and its ratio value can also be used as a set of training samples. Wherein x is any number of fields recorded in the history ratio spelling record. Therefore, a training sample set containing a large number of training samples can be obtained according to each live broadcast ratio of the historical record, and the strength prediction model can be trained according to the training sample set, so that the training effect of the strength prediction model is ensured.

In some embodiments, the history spelling feature of the anchor may be extracted according to the history spelling record of the anchor, and the strength prediction model may be trained. Therefore, useful information of the strength prediction model can be extracted from the history comparison records, and the training effect of the strength prediction model is ensured. Specifically, referring to fig. 10, step S310 may include:

step S311: and acquiring the historical comparison characteristic of the target anchor to be matched, wherein the target anchor to be matched is any anchor in the plurality of anchors to be matched.

Step S312: and acquiring the anchor identification of the anchor to be matched with the target and the matching start time of live broadcast and link matching.

In some embodiments, one anchor may be arbitrarily selected from the plurality of anchors to be matched in the matching pool as a target anchor to be matched to predict the actual force reference value, and then one anchor is selected again to predict the actual force reference value, so that the plurality of anchors to be matched in the matching pool are traversed to predict the actual force reference value.

The history comparison feature, the anchor identifier, and the matching start time may be obtained by referring to the content of the foregoing embodiment, which is not described herein.

Step S313: and performing cross feature operation on the anchor identifier and the matching start time to obtain a second combined feature.

Because the specific collage forces of the anchor under different live broadcast times have a larger gap, in some embodiments, features such as anchor identification and PK time can be introduced into a strength prediction model for prediction, and the model prediction effect is improved. Specifically, the anchor identifier and the matching start time may be input into the cross layer to perform a cross feature operation, so as to obtain a second combined feature. The multi-dimensional characteristics and the relevance characteristics of the anchor and the matching starting time are combined by performing cross characteristic operation, namely nonlinearity is added to the strength prediction model, and the prediction effect of the model is improved.

In some embodiments, referring to fig. 11, step S313 may include:

step S3131: and carrying out Cartesian product operation on the anchor identifier and the matching start time, and obtaining a Cartesian product result.

Step S3132: and Hash conversion is carried out on the Cartesian product result, and the Hash result is obtained as a second combination characteristic.

The operation of the cross feature may participate in the content of the foregoing embodiments, which is not described herein.

Step S314: the historical specific spelling characteristic is input into a linear regression model, and the historical specific spelling characteristic and the second combined characteristic are input into a deep neural network model.

In some embodiments, the above-described strength prediction model may be composed of a linear regression model and a deep neural network model, so that the strength prediction model obtained by training can obtain the memory (modeling) capability of the linear regression model and the generalization (generalization) capability of the DNN model at the same time. Wherein, the memory (memory) discovers the relativity between the history comparing characteristics from the history comparing records. Generalization (generalization), i.e., the transfer of relevance, may find new feature combinations that occur little or no in the history of the comparison. That is, the deep neural network model may explore the high-order information and cross information of features.

In some embodiments, the history specific spelling feature may be input into the linear regression model for prediction, and the history specific spelling feature and the second combined feature may be input into the deep neural network model for prediction, that is, the deep neural network model increases the cross feature between the anchor identifier and the matching start time relative to the linear regression model, so as to consider the situation that the specific spelling force of the anchor is different in different live broadcast time, improve the feature dimension, and ensure the training effect of the model.

Wherein the form of the linear regression model in the real force prediction model may be y=wx+b. Where y is the prediction, x is the input vector of d features, which may be the history of the anchor, w is the model parameters, and b is the bias. The deep neural network model in the strength prediction model may include multiple hidden layers. The multi-layer hidden layer is used for carrying out multi-layer abstraction on the input characteristics to obtain the high-order information of the characteristics, so that different types of data can be divided better in a linear manner. As one approach, the deep neural network model in the real-force prediction model may include 3 hidden layers, where the 3 hidden layer dimensions may be 128, 32, 4, respectively. Thus, the history comparison feature and the second combination feature can be input into a 3-layer hidden layer through an input layer to perform deep learning prediction.

In some embodiments, since different data may have different data dimensions, to facilitate subsequent model training, before the history spell feature and the second combined feature are input into the deep neural network model, the history spell feature, the anchor identifier, and the matching start time may be input into the embedding layer for feature embedding to convert (or reduce dimensions) all features into a feature representation (vector) of a fixed size. And then inputting the embedded anchor identifier and the matching start time into a cross layer for cross feature operation to obtain a second combined feature, so that the embedded history comparison feature and the second combined feature are input into a deep neural network model for prediction.

In some embodiments, since the dimension of the combined feature after the cross feature operation is too high, the second combined feature after the cross feature operation needs to be input to the embedding layer again for feature embedding, so as to perform data dimension reduction (for example, dimension reduction to 10). I.e. an embedded layer is added between the cross layer and the hidden layer.

Step S315: and combining the output of the linear regression model with the output of the deep neural network model to obtain a predicted strength reference value of the anchor to be matched.

In some embodiments, the output of the offline regression model and the output of the deep neural network model may be combined to obtain the predicted strength reference value of the anchor to be matched.

The present embodiment is exemplarily described below with reference to fig. 12:

referring to fig. 12, fig. 12 is a schematic diagram illustrating a data flow of a real force prediction model according to an embodiment of the application. As shown in fig. 12, the real force prediction model includes a wide model part and a deep model part. The window model part can directly input the history comparison feature of the anchor, and output the window model. The deep model part can input the history specific spelling characteristic, the anchor mark and the matching start time into an embedding layer for characteristic embedding, then input the embedded anchor mark and the matching start time into a cross layer for cross combination to obtain a second combined characteristic, and then send the second combined characteristic and the embedded history specific spelling characteristic into a 3-layer neural network (the characteristic dimensions of the 3-layer hidden layer are 128, 32 and 4 respectively) for deep learning, so that the output of the deep model part can be obtained. The output of the wide model part and the output of the deep model part are combined, and finally, the prediction comparison value of the anchor is output.

Step S320: and obtaining the anchor to be matched corresponding to the predicted strength reference value with the largest value as the target anchor with the highest strength reference value.

In some embodiments, after obtaining the predicted actual force reference value of each of the plurality of to-be-matched anchor based on the pre-trained actual force prediction model, the predicted actual force reference values of each to-be-matched anchor may be ranked according to the order of the predicted actual force reference values from large to small, so that the predicted actual force reference values may be ranked into the first corresponding to-be-matched anchor as the target anchor with the highest actual force reference value. In some embodiments, the anchor to be matched corresponding to the predicted strength reference value with the largest value may be directly obtained from multiple anchors to be matched without ordering, and the anchor to be matched is used as the target anchor with the highest strength reference value.

In conclusion, through the fact that the anchor identification and the matching start time feature are introduced through pre-training, the fact reference value of each anchor to be matched can be accurately predicted, so that the final matching result can be more approximately optimized when the bundle searching is conducted based on the fact reference value accurately predicted, and the overall live broadcasting effect can be approximately optimized.

According to the live broadcast and even wheat matching method provided by the embodiment of the application, after the plurality of to-be-matched anchors in the matching pool are obtained, the comparison combination set between the plurality of to-be-matched anchors is obtained, the comparison combination set comprises at least one comparison combination, the comparison combination is composed of two different to-be-matched anchors in the plurality of to-be-matched anchors, so that the predicted total comparison value of each comparison combination in the comparison combination set is obtained based on a pre-trained comparison prediction model, and therefore, a plurality of comparison combinations with the largest sum of the predicted total comparison values can be determined from the comparison combination set according to the predicted total comparison value and a cluster search algorithm to serve as live broadcast and even wheat matching results of the plurality of to-be-matched anchors, wherein the same to-be-matched anchors do not exist in the plurality of comparison combinations. According to the application, the total ratio spelling value of various ratio spelling combinations can be predicted more accurately through the pre-trained ratio spelling prediction model, so that a plurality of anchor to be matched in a matching pool can be reasonably matched according to the total ratio spelling value and the cluster search algorithm, and the final matching result is approximately optimized. Thus comprehensively considering the whole live broadcast effect, improving the retention rate of each main broadcasting live broadcast room and promoting the conversion rate and the success rate of the live broadcast electronic commerce.

Referring to fig. 13, fig. 13 shows a block diagram of a live link matching device 400 according to an embodiment of the present application, where the live link matching device 400 is applied to an electronic device. The live-broadcast wheat-linking matching device 400 includes: the anchor acquisition module 410, the combination acquisition module 420, the prediction acquisition module 430, and the result determination module 440. The anchor obtaining module 410 is configured to obtain a plurality of anchor to be matched in the matching pool; the combination obtaining module 420 is configured to obtain a comparison combination set between the multiple anchor to be matched, where the comparison combination set includes at least one comparison combination, and the comparison combination is formed by two different anchor to be matched in the multiple anchor to be matched; the prediction obtaining module 430 is configured to obtain a predicted total spell value of each spell combination in the spell combination set based on a pre-trained spell prediction model, where the spell prediction model is obtained by training a neural network model according to a training sample, and the training sample includes a spell combination sample and a total spell value sample corresponding to the spell combination sample; the result determining module 440 is configured to determine, from the set of ratio combinations, a plurality of ratio combinations with the largest sum of the predicted total ratio values as a matching result of the live-broadcast even-wheat ratio of the plurality of hosts to be matched, where the plurality of ratio combinations do not have the same host to be matched.

In some embodiments, the prediction acquisition module 430 may include: a feature acquisition unit and a feature input unit. The feature acquisition unit is used for acquiring a first historical comparison feature of a first anchor in a current comparison combination and a second historical comparison feature of a second anchor in the current comparison combination, wherein the current comparison combination is any comparison combination in the comparison combination set; the feature input unit is used for inputting the first historical comparison feature and the second historical comparison feature into a pre-trained comparison prediction model to obtain a predicted total comparison value corresponding to the current comparison combination output by the comparison prediction model, the comparison prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a historical comparison feature sample corresponding to the comparison combination sample and a total comparison value sample corresponding to the historical comparison feature sample.

In some embodiments, the feature input unit may be specifically configured to: acquiring a first anchor identifier of the first anchor and a second anchor identifier of the second anchor; acquiring the matching start time of live broadcast continuous wheat comparison spelling; performing cross feature operation on the first anchor identifier, the second anchor identifier, the matching start time, the first history comparison feature and the second history comparison feature to obtain a first combination feature; and inputting the first combined characteristic into a pre-trained comparison prediction model.

In some embodiments, the result determination module 440 may be specifically configured to: and determining a plurality of comparison combinations with the largest sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values and a cluster search algorithm, and taking the comparison combinations as matching results of the live-broadcast and wheat-linking comparison of the plurality of the hosts to be matched.

Further, in some embodiments, the result determining module 440 may include: from the comparison combination set, according to the order of the predicted total comparison value from large to small, the first i comparison combinations corresponding to a target anchor are obtained and used as current i optimal solution sequences, wherein i is the bundling width, the target anchor is the highest in the anchor to be matched, and the actual force reference value is used for representing the actual nutrient of the anchor to be matched; obtaining the first i ratio combining combinations corresponding to the appointed anchor with the highest strength reference value in the residual ratio combining sets according to the sequence from the big to the small of the predicted total ratio combining value, wherein the residual ratio combining sets are the ratio combining sets of the to-be-matched anchor in the target sequences, the residual to-be-matched anchor is the residual anchor except for the to-be-matched anchor in the target sequences, and the target sequences are any one sequence of the i optimal solution sequences; splicing the first i comparison and splicing combinations corresponding to the designated anchor with the target sequence respectively to obtain i combination sequences corresponding to the target sequence; according to the sum of the predicted total comparison values of the i combined sequences corresponding to each sequence in the i optimal solution sequences, the first i combined sequences are obtained in the order from large to small to be used as new i optimal solution sequences; repeating the steps from the rest ratio combining to the first i ratio combining corresponding to the appointed anchor with the highest strength reference value in the rest to-be-matched anchor according to the sequence from the big to the small of the predicted total ratio combining value, and obtaining the first i combined sequences according to the sum of the predicted total ratio combining values of the i combined sequences corresponding to each sequence in the i optimal solution sequences as new i optimal solution sequences until the i optimal solution sequences contain the plurality of anchor to be matched or only do not contain one anchor in the plurality of anchor to be matched, thereby obtaining final i optimal solution sequences; and obtaining a sequence with the maximum sum of the predicted total comparison values from the final i optimal solution sequences, and taking the sequence as a matching result of the live-broadcast continuous-wheat comparison of the plurality of main broadcasting to be matched.

Further, in some embodiments, the live link matching apparatus 400 may further include: and the strength prediction module and the maximum value acquisition module. The system comprises a power prediction module, a neural network model training module and a power prediction module, wherein the power prediction module is used for acquiring a predicted power reference value of each anchor to be matched in the plurality of anchors to be matched based on a pre-trained power prediction model, and the power prediction model is obtained by training the neural network model according to an anchor sample and a power reference value sample corresponding to the anchor sample; the maximum value acquisition module is used for acquiring the anchor to be matched corresponding to the predicted actual force reference value with the maximum value as the target anchor with the highest actual force reference value.

Further, in some embodiments, the force prediction module may include: the device comprises a target feature acquisition unit, an identification acquisition unit, a combined feature acquisition unit, a combined input unit and an output merging unit. The target feature acquisition unit is used for acquiring the history comparison feature of a target anchor to be matched, wherein the target anchor to be matched is any anchor in the plurality of anchors to be matched; the identification acquisition unit is used for acquiring the anchor identification of the anchor to be matched with the target and the matching start time of live broadcast and link matching; the combined feature acquisition unit is used for carrying out cross feature operation on the anchor identifier and the matching start time to obtain a second combined feature; the combination input unit is used for inputting the history comparison feature into a linear regression model, and inputting the history comparison feature and the second combination feature into a deep neural network model; and the output merging unit is used for merging the output of the linear regression model and the output of the deep neural network model to obtain a prediction strength reference value of the anchor to be matched with the target.

Further, in some embodiments, the combined feature acquisition unit may be specifically configured to: carrying out Cartesian product operation on the anchor identifier and the matching start time, and obtaining a Cartesian product result; and Hash conversion is carried out on the Cartesian product result, and the Hash result is obtained as a second combination characteristic.

The live-broadcast continuous-wheat matching device provided by the embodiment of the application is used for realizing the corresponding live-broadcast continuous-wheat matching method in the embodiment of the method, has the beneficial effects of the corresponding embodiment of the method, and is not repeated herein.

It will be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working process of the apparatus and modules described above may refer to the corresponding process in the foregoing method embodiment, which is not repeated herein.

In the several embodiments provided by the present application, the illustrated or discussed coupling or direct coupling or communication connection of the modules to each other may be through some interfaces, indirect coupling or communication connection of devices or modules, electrical, mechanical, or other forms.

In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist alone physically, or two or more modules may be integrated into one module. The integrated modules may be implemented in hardware or in software functional modules.

Referring to fig. 14, fig. 14 is a block diagram illustrating a structure of an electronic device according to an embodiment of the application. The electronic device 700 may be an electronic device such as a server capable of running applications. The electronic device 700 of the present application may include one or more of the following components: a processor 710, a memory 720, and one or more application programs, wherein the one or more application programs may be stored in the memory 720 and configured to be executed by the one or more processors 710, the one or more program(s) configured to perform the method as described in the foregoing method embodiments.

Processor 710 may include one or more processing cores. The processor 710 utilizes various interfaces and lines to connect various portions of the overall electronic device 700, perform various functions of the electronic device 700, and process data by executing or executing instructions, programs, code sets, or instruction sets stored in the memory 720, and invoking data stored in the memory 720. Alternatively, the processor 710 may be implemented in hardware in at least one of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), programmable logic array (Programmable Logic Array, PLA). The processor 710 may integrate one or a combination of several of a central processing unit (Central Processing Unit, CPU), an image processor (Graphics Processing Unit, GPU), and a modem, etc. The CPU mainly processes an operating system, a user interface, an application program and the like; the GPU is used for being responsible for rendering and drawing of display content; the modem is used to handle wireless communications. It will be appreciated that the modem may not be integrated into the processor 710 and may be implemented solely by a single communication chip.

The Memory 720 may include a random access Memory (Random Access Memory, RAM) or a Read-Only Memory (Read-Only Memory). Memory 720 may be used to store instructions, programs, code, sets of codes, or sets of instructions. The memory 720 may include a stored program area and a stored data area, wherein the stored program area may store instructions for implementing an operating system, instructions for implementing at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the various method embodiments described below, etc. The storage data area may also store data created by the electronic device 700 in use, and the like.

It will be appreciated by those skilled in the art that the structure shown in fig. 14 is merely a block diagram of a portion of the structure associated with the present inventive arrangements and is not limiting of the electronic device to which the present inventive arrangements are applied, and that a particular electronic device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.

In summary, according to the live-broadcast and even-wheat matching method, the device and the electronic equipment provided by the embodiment of the application, by acquiring the comparison combination set among a plurality of to-be-matched anchors, the comparison combination set comprises at least one comparison combination, the comparison combination is composed of two different to-be-matched anchors among the plurality of to-be-matched anchors, so that the predicted total comparison value of each comparison combination in the comparison combination set is acquired based on a pre-trained comparison prediction model, and therefore, a plurality of comparison combinations with the largest sum of the predicted total comparison values can be determined from the comparison combination set according to the predicted total comparison value to serve as the matching result of the live-broadcast and even-wheat comparison of the plurality of to-be-matched anchors, wherein the plurality of comparison combinations do not have the same to-be-matched anchor. According to the application, the total ratio spelling value of various ratio spelling combinations can be predicted more accurately through the pre-trained ratio spelling prediction model, so that a plurality of anchor to be matched in the matching pool can be reasonably matched according to the total ratio spelling value. By combining a plurality of comparison and spelling with the largest sum of the predicted total comparison and spelling values, the scheme that the total comparison and spelling values of the plurality of comparison and spelling combinations matched due to unreasonable pairing cannot be too high can be eliminated as the matching result of the live broadcast and wheat connection and comparison and spelling of the matching pool, the problem that the live broadcast level difference of two paired main broadcasting parties is large is avoided, meanwhile, the intensity of live broadcast and wheat connection PK can be improved, the live broadcast and comparison and spelling effect is improved, and the retention rate of each main broadcasting and live broadcasting room is further improved.

Referring to fig. 15, a block diagram of a computer readable storage medium according to an embodiment of the application is shown. The computer readable storage medium 800 has stored therein program code that can be invoked by a processor to perform the methods described in the method embodiments described above.

The computer readable storage medium 800 may be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read only memory), an EPROM, a hard disk, or a ROM. Optionally, the computer readable storage medium 800 comprises a non-transitory computer readable medium (non-transitory computer-readable storage medium). The computer readable storage medium 800 has storage space for program code 810 that performs any of the method steps described above. The program code can be read from or written to one or more computer program products. Program code 810 may be compressed, for example, in a suitable form.

In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, the different embodiments or examples described in this specification and the features of the different embodiments or examples may be combined and combined by those skilled in the art without contradiction.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and are not limiting; although the application has been described in detail with reference to the foregoing embodiments, it will be appreciated by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not drive the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A live-broadcast continuous-wheat matching method, characterized in that the method comprises the following steps:

acquiring a plurality of anchor to be matched in a matching pool;

acquiring a comparison combination set among the plurality of anchor to be matched, wherein the comparison combination set comprises at least one comparison combination which is composed of two different anchor to be matched in the plurality of anchor to be matched;

obtaining a predicted total spell value for each of the set of spell combinations based on a pre-trained spell prediction model, the obtaining the predicted total spell value for each of the set of spell combinations based on the pre-trained spell prediction model comprising: acquiring a first historical comparison feature of a first anchor in a current comparison combination and a second historical comparison feature of a second anchor in the current comparison combination, wherein the current comparison combination is any comparison combination in the comparison combination set;

Inputting the first historical comparison feature and the second historical comparison feature into a pre-trained comparison prediction model to obtain a prediction total comparison value corresponding to the current comparison combination output by the comparison prediction model, wherein the comparison prediction model is obtained by training a neural network model according to a training sample, and the training sample comprises a historical comparison feature sample corresponding to the comparison combination sample and a total comparison value sample corresponding to the historical comparison feature sample; the inputting the first historical specific spelling characteristic and the second historical specific spelling characteristic into a pre-trained specific spelling prediction model comprises the following steps: acquiring a first anchor identifier of the first anchor and a second anchor identifier of the second anchor; acquiring the matching start time of live broadcast continuous wheat comparison spelling; performing cross feature operation on the first anchor identifier, the second anchor identifier, the matching start time, the first history comparison feature and the second history comparison feature to obtain a first combination feature; inputting the first combined characteristic into a pre-trained comparison prediction model;

and determining a plurality of comparison combinations with the maximum sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values, wherein the comparison combinations are used as the matching results of the live-broadcast continuous-wheat comparison of the plurality of the hosts to be matched, and the same hosts to be matched do not exist in the comparison combinations.

2. The method according to claim 1, wherein the determining, from the set of ratio combinations, a plurality of ratio combinations with the largest sum of the predicted total ratio values as the matching result of the live link ratio of the plurality of hosts to be matched includes:

and determining a plurality of comparison combinations with the largest sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values and a cluster search algorithm, and taking the comparison combinations as matching results of the live-broadcast and wheat-linking comparison of the plurality of the hosts to be matched.

3. The method according to claim 2, wherein the determining, from the set of ratio combinations, a plurality of ratio combinations with the largest sum of the predicted total ratio values as the matching result of the live link ratio of the plurality of hosts to be matched according to the predicted total ratio values and a bundle search algorithm includes:

from the comparison combination set, according to the order of the predicted total comparison value from large to small, the first i comparison combinations corresponding to a target anchor are obtained and used as current i optimal solution sequences, wherein i is the bundling width, the target anchor is the highest in the anchor to be matched, and the actual force reference value is used for representing the actual nutrient of the anchor to be matched;

The method comprises the steps that from remaining ratio combining, according to the order of the predicted total ratio combining value from big to small, the first i ratio combining corresponding to the appointed anchor with the highest strength reference value in the remaining to-be-matched anchor is obtained, wherein the remaining ratio combining is the ratio combining of the to-be-matched anchor in the ratio combining set, which does not contain the target sequences, the remaining to-be-matched anchor is the remaining anchor except the to-be-matched anchor in the target sequences, and the target sequences are any sequence in the i optimal solution sequences;

splicing the first i comparison and splicing combinations corresponding to the designated anchor with the target sequence respectively to obtain i combination sequences corresponding to the target sequence;

according to the sum of the predicted total comparison values of the i combined sequences corresponding to each sequence in the i optimal solution sequences, the first i combined sequences are obtained in the order from large to small to be used as new i optimal solution sequences;

repeating the steps from the rest ratio combining to the first i ratio combining corresponding to the appointed anchor with the highest strength reference value in the rest anchor to be matched according to the sequence from the big to the small of the predicted total ratio combining value, and obtaining the first i combined sequences as new i optimal solution sequences according to the sum of the predicted total ratio combining values of the i combined sequences corresponding to each sequence in the i optimal solution sequences, until the i optimal solution sequences contain the plurality of anchors to be matched or only do not contain one anchor of the plurality of anchors to be matched, so as to obtain final i optimal solution sequences;

And obtaining a sequence with the maximum sum of the predicted total comparison values from the final i optimal solution sequences, and taking the sequence as a matching result of the live-broadcast continuous-wheat comparison of the plurality of main broadcasting to be matched.

4. The method of claim 3, wherein before the first i ratio combinations corresponding to the target anchor are obtained from the set of ratio combinations in order of the predicted total ratio values from greater to lesser, the method further comprises:

acquiring a predicted strength reference value of each anchor to be matched in the plurality of anchors to be matched based on a pre-trained strength prediction model, wherein the strength prediction model is obtained by training a neural network model according to an anchor sample and a strength reference value sample corresponding to the anchor sample;

and obtaining the anchor to be matched corresponding to the predicted strength reference value with the largest value as the target anchor with the highest strength reference value.

5. The method of claim 4, wherein the strength prediction model is composed of a linear regression model and a deep neural network model, wherein the obtaining the predicted strength reference value for each of the plurality of hosts to be matched based on the pre-trained strength prediction model comprises:

Acquiring historical comparison characteristics of a target anchor to be matched, wherein the target anchor to be matched is any anchor in the plurality of anchors to be matched;

acquiring a main broadcasting identification of the main broadcasting to be matched with the target and matching start time of live broadcasting and wheat connecting comparison;

performing cross feature operation on the anchor identifier and the matching start time to obtain a second combined feature;

inputting the history spelling characteristic into a linear regression model, and inputting the history spelling characteristic and the second combination characteristic into a deep neural network model;

and combining the output of the linear regression model with the output of the deep neural network model to obtain a predicted strength reference value of the anchor to be matched.

6. The method of claim 5, wherein the intersecting the anchor identifier with the matching start time to obtain a second combined feature comprises:

carrying out Cartesian product operation on the anchor identifier and the matching start time, and obtaining a Cartesian product result;

and Hash conversion is carried out on the Cartesian product result, and the Hash result is obtained as a second combination characteristic.

7. A live-feed wheat-linking matching device, characterized in that the device comprises:

The anchor acquisition module is used for acquiring a plurality of anchors to be matched in the matching pool;

the combination acquisition module is used for acquiring a comparison combination set among the plurality of anchor to be matched, wherein the comparison combination set comprises at least one comparison combination which is formed by two different anchor to be matched in the plurality of anchor to be matched;

the prediction obtaining module is configured to obtain a predicted total ratio match value of each ratio match combination in the ratio match combination set based on a pre-trained ratio match prediction model, where the obtaining of the predicted total ratio match value of each ratio match combination in the ratio match combination set based on the pre-trained ratio match prediction model includes: acquiring a first historical comparison feature of a first anchor in a current comparison combination and a second historical comparison feature of a second anchor in the current comparison combination, wherein the current comparison combination is any comparison combination in the comparison combination set;

And the result determining module is used for determining a plurality of comparison combinations with the largest sum of the prediction total comparison values from the comparison combination set according to the prediction total comparison values, and the comparison combinations are used as the matching results of the live-broadcast continuous-cast comparison of the plurality of the hosts to be matched, wherein the same hosts to be matched do not exist in the plurality of comparison combinations.

8. An electronic device, comprising:

one or more processors;

a memory;

one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications configured to perform the method of any of claims 1-6.

9. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a program code, which is callable by a processor for executing the method according to any one of claims 1-6.