Live broadcast recommendation method and device, equipment and storage medium

Abstract

The embodiment of the invention relates to a live broadcast recommendation method and device, equipment and a storage medium. The method includes: determining a plurality of traffic feature sequences of a live stream corresponding to a plurality of interactive operations, the plurality of interactive operations being initiated by a broadcast watching user in the live stream, and in each of the plurality of traffic feature sequences, selecting a plurality of traffic feature sequences corresponding to the plurality of interactive operations; the different traffic features represent feature information corresponding to the interactive operation in different historical time slices of the live broadcast stream; at least based on the plurality of traffic feature sequences, determining predicted traffic changes, corresponding to the plurality of interactive operations, of the live broadcast stream in the target time slice; and determining the recommendation degree of the live broadcast stream for the user group at least based on the recommendation request of the user group and the predicted flow change of the live broadcast stream corresponding to the plurality of interactive operations in the target time slice.

Description

Technical Field

[0001] Example embodiments of the present disclosure generally relate to the field of computer technology, and more specifically, to a live broadcast recommendation method, apparatus, device, and storage medium. Background Art

[0002] Live streaming is a form of communication that uses Internet technology to simultaneously produce and disseminate content, allowing viewers to watch and interact in real time. In live streaming services, live content is recommended to users based on their needs. However, traditional live streaming recommendation schemes generally recommend live streaming based on the audience's perspective, which can lead to some high-quality content being underestimated during the recommendation process. Summary of the invention

[0003] In a first aspect of the present disclosure, a live broadcast recommendation method is provided. The method includes: determining multiple traffic feature sequences corresponding to multiple interactive operations of a live broadcast stream, wherein the multiple interactive operations are initiated by a viewing user in the live broadcast stream, and in each of the multiple traffic feature sequences, different traffic features represent feature information corresponding to the interactive operations in different historical time slices of the live broadcast stream; determining predicted traffic changes corresponding to each of the multiple interactive operations of the live broadcast stream in a target time slice based at least on the multiple traffic feature sequences; and determining the recommendation degree of the live broadcast stream for the user group based at least on the recommendation request of the user group and the predicted traffic changes corresponding to each of the multiple interactive operations of the live broadcast stream in the target time slice.

[0004] In a second aspect of the present disclosure, a device for live broadcast recommendation is provided. The device includes: a traffic feature sequence determination module, configured to determine multiple traffic feature sequences corresponding to multiple interactive operations of a live broadcast stream, wherein the multiple interactive operations are initiated by a viewing user in the live broadcast stream, and in each of the multiple traffic feature sequences, different traffic features represent feature information corresponding to the interactive operations in different historical time slices of the live broadcast stream; a predicted traffic change determination module, configured to determine the predicted traffic changes corresponding to each of the multiple interactive operations of the live broadcast stream in a target time slice based on at least the multiple traffic feature sequences; and a recommendation degree determination module, configured to determine the recommendation degree of the live broadcast stream for the user group based on at least the recommendation request of the user group and the predicted traffic changes corresponding to each of the multiple interactive operations of the live broadcast stream in the target time slice.

[0005] In a third aspect of the present disclosure, an electronic device is provided. The device includes at least one processor; and at least one memory, the at least one memory is coupled to the at least one processor and stores instructions for execution by the at least one processor. When the instructions are executed by the at least one processor, the device executes the method of the first aspect.

[0006] In a fourth aspect of the present disclosure, a computer-readable storage medium is provided, wherein computer-executable instructions are stored on the computer-readable storage medium, and the computer-executable instructions can be executed by a processor to implement the method of the first aspect.

[0007] In a fifth aspect of the present disclosure, a computer program product is provided, which includes computer executable instructions, and when the instructions are executed by a processor, the method according to the first aspect of the present disclosure is implemented.

[0008] It should be understood that the contents described in this content section are not intended to limit the key features or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0010] Figure 1 A schematic diagram showing an example environment in which embodiments of the present disclosure can be implemented;

[0011] Figure 2 A flowchart showing an example process of a live broadcast recommendation method according to some embodiments of the present disclosure;

[0012] Figure 3 A schematic diagram showing an example architecture of a content recommendation system according to some embodiments of the present disclosure;

[0013] Figure 4 A schematic diagram showing an example of determining a flow change tag according to some embodiments of the present disclosure;

[0014] Figure 5 A schematic diagram showing an example of determining a predicted flow change according to some embodiments of the present disclosure;

[0015] Figure 6 A schematic diagram showing an example of adjusting a recommendation ranking according to some embodiments of the present disclosure;

[0016] Figure 7 A schematic structural block diagram of an apparatus for live broadcast recommendation according to some embodiments of the present disclosure is shown; and

[0017] Figure 8 A block diagram of an electronic device is shown in which one or more embodiments of the present disclosure may be implemented. DETAILED DESCRIPTION

[0018] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0019] It should be noted that the titles of any sections / subsections provided herein are not restrictive. Various embodiments are described throughout this article, and any type of embodiment may be included under any section / subsection. In addition, the embodiments described in any section / subsection may be combined in any manner with any other embodiments described in the same section / subsection and / or different sections / subsections.

[0020] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The term "some embodiments" should be understood as "at least some embodiments". Other explicit and implicit definitions may be included below. The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may be included below.

[0021] The embodiments of the present disclosure may involve user data, data acquisition and / or use, etc. These aspects are subject to the corresponding laws, regulations and relevant provisions. In the embodiments of the present disclosure, all data collection, acquisition, processing, processing, forwarding, use, etc. are carried out on the premise that the user knows and confirms. Accordingly, when implementing each embodiment of the present disclosure, the type, scope of use, usage scenario, etc. of the data or information that may be involved should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with the relevant laws and regulations. The specific notification and / or authorization method can vary according to the actual situation and application scenario, and the scope of the present disclosure is not limited in this respect.

[0022] In this specification and the embodiments, if personal information processing is involved, it will be processed on the premise of having a legal basis (such as obtaining the consent of the subject of personal information, or it is necessary to perform a contract, etc.), and will only be processed within the scope of regulations or agreements. If a user refuses to process personal information other than the necessary information for basic functions, it will not affect the user's use of basic functions.

[0023] As briefly described above, traditional live broadcast recommendation schemes generally recommend live broadcasts based on the audience's perspective, which will cause some high-quality content to be underestimated in the recommendation process. Specifically, compared with traditional videos, live broadcasts are an interactive medium between anchors (also known as live broadcast suppliers) and viewers (also known as live broadcast demanders). The production process and extraction process of live broadcasts exist simultaneously and influence each other. For example, assuming that the anchor successively conducts activities such as "chatting with the audience" and "dance performance", the traffic efficiency of different activities will show significant differences. Traffic efficiency can be an indicator used to measure the live broadcast content in attracting viewers to watch and promoting audience interaction. Traffic efficiency can be determined by the number of comments and / or likes from the audience. Compared with "chatting with the audience", the number of comments and / or likes from the audience is often higher when the anchor is performing "dance performance". This means that high-quality live broadcast content can attract the audience to actively participate in the interaction, thereby improving traffic efficiency. Such positive feedback will then affect the anchor's live broadcast method, prompting the anchor to be more inclined to provide high-quality content in the future (such live broadcast content is also referred to as potential high-quality content below).

[0024] Traditional live broadcast recommendation schemes generally recommend live broadcasts that viewers may be interested in based on their historical viewing data. Such live broadcast recommendation schemes do not take into account the interaction between viewers and live broadcasts, resulting in the aforementioned potential high-quality content being underestimated in the recommendation process.

[0025] In view of this, an embodiment of the present disclosure provides a live broadcast recommendation scheme. According to the scheme, firstly, a plurality of traffic feature sequences corresponding to a plurality of interactive operations of a live broadcast stream are determined, wherein the plurality of interactive operations are initiated by viewing users in the live broadcast stream, and in each of the plurality of traffic feature sequences, different traffic features represent feature information corresponding to the interactive operations in different historical time slices of the live broadcast stream. Then, based at least on the plurality of traffic feature sequences, the predicted traffic changes corresponding to each of the plurality of interactive operations of the live broadcast stream in the target time slice are determined (for ease of discussion, these predicted traffic changes will hereinafter also be collectively or individually referred to as interactive traffic changes). Subsequently, based at least on the recommendation request of the user group and the predicted traffic changes corresponding to each of the plurality of interactive operations of the live broadcast stream in the target time slice, the recommendation degree of the live broadcast stream for the user group is determined.

[0026] It will be more clearly understood through the following description that the live broadcast recommendation scheme proposed in the present disclosure is based on the live broadcast stream (also referred to as the transmission carrier of the live broadcast content in the live broadcast room). Such a live broadcast recommendation scheme can comprehensively capture the interactive operations (such as comments and / or likes, etc.) of the viewing users (also referred to as the audience) during the live broadcast process based on the perspective of the live broadcast room. Furthermore, the scheme of the present disclosure extracts features of the interactive operations in the live broadcast stream based on historical time slices (for example, extracts feature information based on minute-level historical time slices) and constructs corresponding traffic feature sequences. Such traffic feature sequences can reflect the traffic efficiency of the live broadcast content in real time in the recent period of time. Based on such traffic feature sequences, the scheme of the present disclosure can accurately predict the next interactive traffic changes of the live broadcast stream (for example, an increase in comment traffic or a decrease in comment traffic), thereby identifying whether the live broadcast stream is the potential high-quality content described above. In the case where the live broadcast stream is the potential high-quality content described above, the embodiments of the present disclosure can improve the recommendation level of the live broadcast stream, thereby avoiding such live broadcast content from being underestimated in the recommendation process.

[0027] In this way, the disclosed solution can provide more active positive feedback to the anchor, thereby optimizing the anchor's live broadcast experience. Active positive feedback is conducive to promoting the emergence of more high-quality content, thereby improving the viewing experience of live broadcast viewers. Therefore, the disclosed solution can simultaneously optimize the bilateral experience of the anchor and the audience, which makes the recommendation process more in line with the characteristics of the live broadcast scene, thereby improving the accuracy of live broadcast recommendations.

[0028] Various example implementations of the solution will be described in detail below in conjunction with the accompanying drawings.

[0029] Figure 1 A schematic diagram of an example environment 100 in which embodiments of the present disclosure can be implemented is shown. In the environment 100, a user 110 may be referred to as a host, a live broadcaster, or a live broadcast provider of a live broadcast room. The user 110 may create and manage a live broadcast room through an associated terminal device 120, thereby providing various live broadcast contents including audio and / or video. It is worth noting that although Figure 1 Only one anchor is shown in the video, but in fact, a live broadcast room can be initiated and managed by multiple anchors to meet the needs of a wider range of users.

[0030] In the environment 100, users 130-1 to 130-N may be referred to as viewers, viewers, viewers, or participants in the live broadcast room, etc., where N is a positive integer. Users 130-1 to 130-N may watch the live broadcast content and participate in the interaction in the live broadcast room through their associated terminal devices 140-1 to 140-N. The terminal devices 140-1 to 140-N may present a live broadcast interface to ensure that viewers can watch the live broadcast clearly and smoothly. For ease of discussion, users 130-1 to 130-N are also collectively or individually referred to as users 130 below, and the corresponding terminal devices are also collectively or individually referred to as terminal devices 140.

[0031] In the environment 100, the terminal device 120 and the terminal device 140 are not only used to present live content, but also can communicate with the content recommendation system 150 through a communication method such as a network. The content recommendation system 150 can be an application, a website, a webpage, and other accessible platforms. The terminal device 120 and the terminal device 140 can be installed with an application for accessing the content recommendation system 150, or the terminal device 120 and the terminal device 140 can access the content recommendation system 150 in any suitable manner.

[0032] The content recommendation system 150 may be configured to recommend live content of one or more hosts to a user group (e.g., users 130-1 to 130-N) based on corresponding strategies. For example, the content recommendation system 150 may recommend live content that may be of interest to the user 130 based on the user's historical viewing data.

[0033] In the environment 100, the terminal device 130 can be any type of mobile terminal, fixed terminal or portable terminal, including a mobile phone, a desktop computer, a laptop computer, a notebook computer, a netbook computer, a tablet computer, a media computer, a multimedia tablet, a personal communication system (PCS) device, a personal navigation device, a personal digital assistant (PDA), an audio / video player, a digital camera / camcorder, a positioning device, a television receiver, a radio broadcast receiver, an e-book device, a gaming device or any combination of the foregoing, including accessories and peripherals of these devices or any combination thereof. In some embodiments, the terminal device 130 can also support any type of interface for the user (such as a "wearable" circuit, etc.).

[0034] In the environment 100, the content recommendation system 150 can be deployed in any type of server-side device. The server-side device can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks, and big data and artificial intelligence platforms. The server-side device can include, for example, a computing system / server, such as a mainframe, an edge computing node, a computing device in a cloud environment, and the like.

[0035] It should be understood that the structure and function of the various elements in the environment 100 are described for exemplary purposes only and do not imply any limitation on the scope of the present disclosure.

[0036] Figure 2 A flowchart of an example process 200 of a live broadcast recommendation method according to some embodiments of the present disclosure is shown. The process 200 may be implemented at the content recommendation system 150.

[0037] Reference Figure 2 In block 210, the content recommendation system 150 determines a plurality of traffic feature sequences corresponding to a plurality of interactive operations of the live stream. The plurality of interactive operations are initiated by a viewing user (e.g., user 130) in the live stream. In the traffic feature sequence corresponding to each interactive operation, different traffic features represent feature information corresponding to the interactive operation in different historical time slices of the live stream.

[0038] As an example, a live stream can be a transmission carrier for live content in a live broadcast room. When the host (e.g., user 110) starts a live broadcast, the live content can be processed into a live stream, and thus transmitted in real time to each viewing user (e.g., terminal device 140). In this way, no matter where the viewing user is, as long as there is a network connection, the host's live content can be viewed instantly. In addition, the live stream also carries the interactive data of the live broadcast room. During the live broadcast, the viewing user can interact with the host through comment operations and / or like operations. These interactive data will also be included in the live stream, and thus transmitted in real time to the host (e.g., terminal device 120) and other viewing users.

[0039] As an example, multiple interactive operations may be various interactions initiated by the viewing user during the live broadcast process, and these interactions may be interactions that can reflect the viewing user's participation and interest in the live broadcast process. As an example, multiple interactive operations may include the viewing user's comment operation and / or like operation, etc. It should be noted that the above is only an exemplary description. According to actual needs, the interactive operations in the embodiments of the present disclosure may also include more operations such as sending virtual gifts and sharing operations, and the embodiments of the present disclosure do not limit this.

[0040] As an example, a time slice may be a number of segments into which the content recommendation system 150 divides the live stream in chronological order. A historical time slice may refer to a time slice of the live stream that is located before the current moment. As an example, a time slice may be a segment in units of minutes. For example, a time slice may be a segment in units of 1 minute / 5 minutes / 10 minutes, etc. It should be noted that the above is only an exemplary description. According to actual needs, a time slice may also be a segment in units of 15 minutes, etc., and the embodiments of the present disclosure do not limit this.

[0041] As an example, the feature information in each historical time slice may be the quantitative data or index of the corresponding interactive operation in the historical time slice. As an example, the feature information corresponding to the comment operation may indicate the amount of comments or the ratio of the amount of comments to the number of users watching the broadcast (also referred to as the comment rate), etc. It should be noted that the above is only an exemplary description. According to actual needs, the feature information may also indicate more content. For example, the feature information corresponding to the attention operation may indicate the amount of attention and the attention rate, etc., and the embodiments of the present disclosure do not limit this. The traffic feature may characterize one or more feature information of the corresponding interactive operation in a single historical time slice. For example, the traffic feature corresponding to the comment may simultaneously characterize the amount of comments and the comment rate, etc. The traffic feature may be used to measure the effect of the corresponding historical time slice in attracting viewers to watch and promoting audience interaction. Therefore, such traffic features may indicate the traffic efficiency of the corresponding historical time slice. In the embodiments of the present disclosure, the traffic efficiency may be defined as the interaction rate between the watching users and the live stream in a single time slice.

[0042] As an example, a traffic feature sequence may refer to a sequence of traffic features corresponding to the same interactive operation in different historical time slices arranged in chronological order. Such a traffic feature sequence can indicate the historical traffic changes of the live stream corresponding to the interactive operation in multiple historical time slices, thereby reflecting the traffic efficiency changes of the live stream in multiple historical time slices. For ease of discussion, the following will refer to multiple traffic changes corresponding to multiple interactive operations (including historical traffic changes and predicted traffic changes) collectively or individually as interactive traffic changes.

[0043] As an example, assuming that the t-3, t-2, and t-1 time slices in the live stream are historical time slices, the traffic feature sequence S corresponding to the comment operation determined based on the t-3, t-2, and t-1 time slices is comment It can be expressed as:

[0044] S comment =[..., ε t-3 , ε t-2 , ε t-1 ];(1)

[0045] Where t is a positive integer, ε t-3 represents the traffic characteristics (e.g., the number of comments) corresponding to the comment operation determined based on the t-3th historical time slice, ε t-2 represents the traffic feature corresponding to the comment operation determined based on the t-2th historical time slice, ε t-1 Represents the traffic features corresponding to the comment operation determined based on the t-1th historical time slice.

[0046] In this way, the embodiments of the present disclosure can build a highly timely traffic feature sequence from the perspective of the live broadcast room based on minute-level (or finer-grained) time slices by aggregating the real-time interactive operations of the live broadcast room. In some embodiments, the length and number of traffic feature sequences can be determined according to actual needs, and the embodiments of the present disclosure do not limit this. As an example, the length of the traffic feature sequence can be set to 15, the number of traffic feature sequences can be set to 16, and so on.

[0047] In block 220 , the content recommendation system 150 determines predicted traffic changes corresponding to each of the multiple interactive operations in the live stream in the target time slice based at least on the multiple traffic feature sequences.

[0048] As an example, the target time slice may refer to the time slice of the current moment of the live stream and / or a period of time after the current moment. As described above, through multiple traffic feature sequences, the traffic efficiency changes of the live stream in multiple historical time slices can be reflected. By determining the predicted traffic changes of the target time slice, the traffic efficiency trend of the live stream in the current time slice or one or more future time slices can be evaluated.

[0049] Figure 3 A schematic diagram showing an example architecture 300 of a content recommendation system 150 according to some embodiments of the present disclosure is shown. Figure 3In some embodiments, the content recommendation system 150 generally includes a live stream processing module 301 and a second machine learning model 302. The live stream processing module 301 is capable of processing a live stream in the form of a data stream and providing a traffic feature sequence to the second machine learning model 302 as its input data. In an embodiment of the present disclosure, the live stream processing module 301 may also be referred to as a data stream engine (SelfFlow). As an example, the live stream processing module 301 may trigger a slice instance service 303 based on a first predetermined period 303 (e.g., 30 seconds or any other appropriate time). The slice instance service 303 may obtain the room identifiers of one or more live broadcast rooms and their basic configuration files. For example, the slice instance service 303 may determine the traffic features by calling (3031) the live broadcast room service 304. Next, the slice instance service 303 may determine the traffic features of each live broadcast room based on the room identifiers of the acquired live broadcast rooms and their basic configuration files, with time slices as the granularity. For example, the slice instance service 303 may determine the traffic features by calling (3032) the feature service 305. Next, the slice instance service 303 may construct multiple traffic feature sequences based on the determined traffic features. Then, the slice instance service 303 determines the predicted traffic changes corresponding to each of the multiple interaction operations by calling (3033) the second machine learning model 302 based on at least the multiple traffic feature sequences. As an example, the slice instance service 303 can call the second machine learning model 302 based on a second predetermined period (e.g., 10 seconds or any other appropriate time) to determine the predicted traffic changes.

[0050] As an example, assuming that the target time slice is the tth time slice and its future x-1 time slices, and the multiple historical time slices are the x time slices before the tth time slice, the process of the second machine learning model 302 determining the predicted traffic change can be expressed by formula (2), where x is a positive integer and x can be set according to actual needs.

[0051]

[0052] in represents the predicted traffic changes corresponding to the comment operation, the follow operation and the like operation in the target time slice, respectively, F(·) represents the algorithm executed by the second machine learning model 302, θ represents the learnable parameters of the second machine learning model 302, and Z represents the input data of the second machine learning model 302. The input data Z at least includes multiple traffic feature sequences determined based on the x time slices before the tth time slice. It should be noted that the above algorithms and parameters for predicting traffic changes are only exemplary. According to actual needs, other algorithms and parameters can also be used to determine the predicted traffic changes.

[0053] As an example, predicting traffic changes It can indicate whether the interaction traffic corresponding to comment operations, follow operations, and like operations will be improved in the target time slice compared to multiple historical time slices. As an example, predicting traffic changes It can be expressed in the form of binary classification, for example, predicting traffic changes "0" indicates that the interactive traffic of the corresponding interactive operation will decrease, and "1" indicates that the interactive traffic of the corresponding interactive operation will increase. It can also be expressed in the form of probability, which is not limited in the embodiments of the present disclosure.

[0054] Continue to refer to Figure 3 In some embodiments, the second machine learning model 302 may be an online machine learning model, and the live stream processing module 301 may construct a traffic feature sequence sample based on the attribution unit 306 (this process may also be referred to as traffic change attribution), and then update the learnable parameter θ of the second machine learning model 302 online through model training.

[0055] Specifically, after determining the traffic characteristics, the slice instance service 303 can store the determined traffic characteristics as the traffic characteristic samples (also referred to as traffic characteristic instances) of the corresponding historical time slice. Then, in response to the trigger associated with the given historical time slice being triggered, the attribution unit 306 can construct a traffic characteristic sequence sample based on the traffic characteristic samples of the given historical time slice and several historical time slices adjacent to the given historical time slice before and after. In addition, the attribution unit 306 can also determine the traffic change labels corresponding to the multiple interactive operations of the given historical time slice based on the traffic characteristic samples of the given historical time slice and several historical time slices adjacent to the given historical time slice before and after. Next, the live stream processing module 301 can update the learnable parameter θ of the second machine learning model 302 through model training based on at least the traffic characteristic sequence samples and the traffic change labels. The traffic change label of the given historical time slice indicates the traffic change of the given historical time slice and several historical time slices located after the given historical time slice in the corresponding interactive operation relative to the historical time slice located before the given historical time slice.

[0056] As an example, the slice instance service 303 may store (3034) the determined traffic feature in the first message queue 307. The attribution unit 306 may extract (3035) the traffic feature sample from the first message queue 307 using the extractor 308 or the like. Subsequently, the extractor 308 may store (3036) the traffic feature sample in the storage unit 309 according to a predetermined data structure, and the storage unit 309 may be, for example, a cache unit. As an example, the predetermined data structure of the traffic feature sample may be a key-value structure, the key in the key-value structure may be the room identifier of the live broadcast room to which the traffic feature belongs, and the value may be the slice identifier of the time slice to which the traffic feature belongs in the time slice list. As an example, the time slices in the time slice list may be arranged in the chronological order of the time slices.

[0057] As an example, the attribution unit 306 may configure (3037) a trigger 3010 for a time slice of the live stream, and the trigger 3010 may send (3038) a trigger indication for a certain time slice to the connector 3011 in the attribution unit 306, and the connector 3011 may, upon receiving the trigger indication, construct a traffic feature sequence sample by merging (3039) traffic feature samples of related historical time slices (e.g., several historical time slices that are adjacent to each other in the same live room). Assuming that the given historical time slice is the t-th time slice, when the t-th time slice is triggered, the connector 3011 may determine the historical time slice that belongs to the same live room as the t-th time slice based on the room identifier, and further determine the traffic feature samples of the t-th time slice and several historical time slices that are adjacent to each other before and after the t-th time slice (e.g., the t-5th time slice to the t-1th time slice and the t+1th time slice to the t+4th time slice) based on the slice identifier. Then, the attribution unit 306 can construct a traffic feature sequence sample based on the determined traffic feature samples. In addition, the attribution unit 306 can delete expired time slices in the time slice list (eg, determined by a predetermined expiration time), thereby releasing storage space of the time slice list.

[0058] As an example, assume that the given historical time slice is the tth time slice, and the several historical time slices before and after the given historical time slice are the t-5th to t-1th time slices and the t+1th to t+4th time slices. Then, the traffic feature samples corresponding to the comment operation in the t-5th to t+4th time slices can be expressed as (e.g. comment rate), i=t-5, t-4, ..., t+4. Traffic change label corresponding to the comment operation at the t-th time slice It can be expressed by formula (3).

[0059]

[0060] Based on a similar approach, the attribution unit 306 can also be based on the traffic feature samples corresponding to the concerned operation in the time slices t-5 to t+4 (e.g., attention rate), determine the traffic change label corresponding to the attention operation at the tth time slice In addition, the attribution unit 306 can also be based on the traffic feature samples corresponding to the like operation in the time slices t-5 to t+4 (e.g. like rate), determine the traffic change label corresponding to the like operation at the tth time slice etc.

[0061] As an example, the flow change label Indicates whether the interactive traffic corresponding to the interactive operation from time slices t to t+4 is improved compared to time slices t-5 to t-1. As an example, the traffic change label It can be expressed in the form of binary classification or probability, which can be determined according to actual needs, and the embodiments of the present disclosure are not limited to this.

[0062] Figure 4 A schematic diagram showing an example 400 of determining a flow change tag according to some embodiments of the present disclosure is shown, referring to Figure 4 , the following takes the binary classification form as an example to determine the flow change label In the live stream 403, assuming that the given historical time slice is the t-th time slice, the time slices adjacent to the t-th time slice are only the t-1-th time slices, and the traffic feature samples are Indicates the comment rate in the corresponding time slice. In the t-1th time slice, suppose there are three viewers 401 and a certain type of interactive operation 402 (such as a comment operation), then the traffic feature sample corresponding to the comment operation is Similarly, in the tth time slice, assuming that there are four viewing users 401 and two interactive operations 402 of the same type (such as comment operations), then the traffic feature sample corresponding to the comment operation is because Therefore, it can be considered that the comment rate of the live stream has increased, and then, the traffic change label corresponding to the comment operation at the tth time slice can be considered to be On the contrary, we can consider that the traffic change label corresponding to the comment operation at the tth time slice is And so on.

[0063] Return to reference Figure 3Once the traffic feature samples are determined, the attribution unit 306 can construct a traffic feature sequence sample F based on these traffic feature samples. t Then, the attribution unit 306 can be based on the traffic feature sequence sample F t and flow change label samples Constructing training samples for the second machine learning model 302 Then, the attribution unit 306 can convert the training sample D t Store (3040) to the second message queue 3012.

[0064] Next, the live stream processing module 301 may update the learnable parameter θ in the second machine learning model 302 through model training based on the training samples in the second message queue 3012. As an example, the sample acquisition unit 3013 in the live stream processing module 301 may extract (3041) the training sample D from the second message queue 3012. t The training unit 3013 in the live stream processing module 301 can be based on the training sample D t Model training is performed to determine the latest learnable parameter θ. Then, the parameter updating unit 3014 in the live stream processing module 301 obtains (3042) the latest learnable parameter θ, and then updates (3043) the second machine learning model 302 based on the obtained learnable parameter θ. As an example, the second machine learning model 302 is based on the training sample D t Determine predicted traffic changes The process can be expressed by formula (4).

[0065]

[0066] In some embodiments, the second machine learning model 302 may be a multi-task learning model, and each task may correspond to a predicted traffic change of an interactive operation 402. Through the multi-task learning model, the embodiments of the present disclosure may share the feature extraction layer of the model to learn common features between multiple tasks, thereby improving overall learning efficiency and performance.

[0067] It should be noted that the above algorithms and parameters are only exemplary and do not constitute a limitation on the embodiments of the present disclosure. According to actual needs, the embodiments of the present disclosure can adopt any appropriate algorithm and parameters, which will not be listed here one by one.

[0068] In some embodiments, the content recommendation system 150 may determine the dependency between each traffic feature in the plurality of traffic feature sequences and other traffic features other than the traffic feature. Then, the content recommendation system 150 extracts the first feature representation of the plurality of traffic feature sequences based on the determined dependency. Subsequently, the content recommendation system 150 determines the predicted traffic changes corresponding to each of the plurality of interactive operations 402 in the target time slice of the live stream 403 based at least on the first feature representation.

[0069] As an example, the dependency between multiple traffic features can be the dependency between multiple traffic features in the same traffic feature sequence, or the dependency between multiple traffic features in different traffic feature sequences. The dependency between multiple traffic features can indicate the correlation between these traffic features. For example, when one traffic feature (such as the comment rate of a comment operation) changes, another traffic feature (such as the like rate of a like operation) also changes accordingly. By analyzing the dependencies between these traffic features, the content recommendation system 150 can better understand the interactions between multiple traffic features, thereby providing an effective basis for subsequent traffic change predictions. As an example, the first feature representation can be a feature vector, feature matrix and / or embedded representation generated by feature encoding, etc. Through the first feature representation, the content recommendation system 150 can convert complex traffic feature sequences into concise and effective feature representations, thereby facilitating subsequent processing and analysis.

[0070] In some embodiments, the dependency relationship between at least one traffic feature and other traffic features other than the traffic feature includes at least a first dependency relationship and / or a second dependency relationship. The first dependency relationship indicates the dependency relationship between traffic features corresponding to different interaction operations 402 in the same historical time slice. The second dependency relationship indicates the dependency relationship between traffic features belonging to different historical time slices in the same traffic feature sequence.

[0071] As an example, within the same time slice of the live stream 403, the user 401 watching the broadcast may perform multiple interactive operations 402, such as commenting and / or giving virtual gifts. The first dependency relationship may indicate the relationship between the traffic features (e.g., comment rate and / or virtual gift giving rate, etc.) corresponding to these interactive operations 402 within the same time slice. For example, the comment rate may show a trend of changing with the virtual gift giving rate. In the embodiment of the present disclosure, the first dependency relationship may also be referred to as the spatial dependency relationship between the traffic features corresponding to the multiple interactive operations 402. The first dependency relationship helps the content recommendation system 150 to more accurately capture the association between different interactive operations 402 of the user 401 watching the broadcast within the same time slice.

[0072] As an example, the traffic characteristics (such as comment rate) corresponding to the same interactive operation 402 may have different performances in different time slices. For example, during the live broadcast, the anchor performs different live broadcast activities in different time slices, and the comment rate may show a trend of changing with the live broadcast activities. In the embodiment of the present disclosure, the second dependency relationship can also be referred to as the time dependency relationship between multiple traffic characteristics corresponding to the same interactive operation 402. The second dependency relationship helps the content recommendation system 150 to more accurately capture the association between the same interactive operation 402 of the viewing user 401 in different time slices.

[0073] It should be noted that the above dependencies between traffic features are only exemplary. According to actual needs, the dependencies between traffic features and other traffic features other than the traffic features may also include more dependencies, which can be determined according to actual needs, and the embodiments of the present disclosure will not be listed one by one here.

[0074] In some embodiments, the content recommendation system 150 may extract multiple graph structure representations corresponding to multiple historical time slices of the live stream 403 from multiple traffic feature sequences based on the determined dependency, wherein for a given graph structure representation corresponding to a given historical time slice, the nodes in the given graph structure representation correspond to the traffic features belonging to the given historical time slice in the multiple traffic feature sequences, and the edge between at least two nodes in the given graph structure representation is determined based on the dependency between the corresponding two traffic features. Then, the content recommendation system 150 may determine the first feature representation through feature encoding based on the dependency between the multiple graph structure representations.

[0075] As an example, for each historical time slice, the content recommendation system 150 can generate a corresponding graph structure representation. In this graph structure, each node can correspond to a traffic feature, and the edges between the nodes represent the dependencies between these traffic features. For example, if the traffic features of a time slice include the number of likes and / or comments, then these traffic features will be used as nodes in the corresponding graph structure representation, and the dependencies between them (such as the comment rate changes with the virtual gift giving rate) are represented by edges. Such a graph structure representation can indicate the spatial dependencies between the multiple traffic features described above.

[0076] Figure 5 A schematic diagram illustrating an example 500 of determining predicted flow changes according to some embodiments of the present disclosure is shown. Figure 5 A plurality of traffic feature sequences 508 are shown, and different traffic feature sequences correspond to different interactive operations, for example Figure 5The multiple traffic feature sequences 508 shown in the figure correspond to comment operations, like operations, and virtual gift giving operations, etc. Figure 5 A plurality of boxes 5010-1, 5010-2, ..., 5010-t-1 are also shown. Assuming that the plurality of historical time slices are from the 1st time slice to the t-1th time slice, the flow feature 509 shown in the box 5010-1 may be the flow feature 509 determined based on the 1st time slice, the flow feature 509 shown in the box 5010-2 may be the flow feature 509 determined based on the 2nd time slice, the flow feature 509 shown in the box 5010-t-1 may be the flow feature 509 determined based on the t-1th time slice, and so on.

[0077] The content recommendation system 150 can use the feature representation extraction module 501 to determine multiple graph structure representations 502-1, 502-2, ..., 502-t-1 corresponding to multiple historical time slices by means such as graph embedding (GraphEmbedding) and / or graph neural network (GNN). For the convenience of discussion, the multiple graph structure representations 502-1, 502-2, ..., 502-t-1 are also collectively or individually referred to as graph structure representations 502 below. As an example, assuming that the multiple historical time slices are the 1st time slice to the t-1th time slice, the graph structure representation 502-1 is the graph structure representation 502 corresponding to the 1st time slice, the graph structure representation 502-2 is the graph structure representation 502 corresponding to the 2nd time slice, and the graph structure representation 502-t-1 is the graph structure representation 502 corresponding to the t-1th time slice. Graph embedding can make similar nodes in the graph structure representation 502 closer in the vector space through a low-dimensional vector space. Graph neural networks can predict unknown information in a graph by learning the features of nodes and edges in the graph structure representation 502.

[0078] Next, the feature representation extraction module 501 can use the first machine learning model 503 (e.g., a Transformer model) to identify the dependencies between the multiple graph structure representations 502, and perform feature encoding to determine the first feature representation. The first feature representation can simultaneously indicate the temporal dependency and spatial dependency between the multiple traffic features described above. In an embodiment of the present disclosure, the feature representation extraction module 501 may also be referred to as a spatiotemporal fusion module.

[0079] In this way, the content recommendation system 150 can use the feature representation extraction module 501 to model the interaction traffic change trend of the interactive operation 402 in the historical time slice at the time slice granularity, thereby deeply exploring the potential spatiotemporal dependencies between multiple traffic features based on the perspective of the live broadcast room.

[0080] In some embodiments, the plurality of graph structure representations 502 are extracted from the plurality of traffic feature sequences using at least graph diffusion convolution. In some embodiments, the dependencies between the plurality of graph structure representations 502 are determined using at least a first machine learning model having a multi-head attention mechanism.

[0081] Continue to refer to Figure 5 The feature representation extraction module 501 can process the plurality of traffic feature sequences 508 into a first matrix 504 of a non-Euclidean structure (or other forms), and the first matrix 504 can also be represented as a matrix X, X∈R N×T , where R represents a real number, N represents the number of traffic feature sequences 508, and T represents the length of the traffic feature sequence 508. As an example, the feature representation extraction module 501 can determine the second matrix 505 based on the first matrix 504, the first learnable parameter 506, the second learnable parameter 507 and the transpose 507 of the first matrix 504. The second matrix 505 can also be represented as an adaptive attention matrix Then, the feature representation extraction module 501 uses graph diffusion convolution to extract the spatial dependency between multiple traffic features 509 from the second matrix 505, thereby obtaining multiple graph structure representations 502.

[0082] As an example, the second matrix 505 (i.e., the adaptive attention matrix ) can be expressed by formula (5).

[0083]

[0084] Among them, T w1 and T w2 denote the first learnable parameter 506 and the second learnable parameter 507, respectively, w1 ∈R T×T , T w2 ∈R T×3 SoftMax(Relu((XT w1 )(XT w2 ) T )) indicates the (Relu((XT w1 )(XT w2 ) T )) for normalization, Relu((XT w1 )(XT w2 ) T ) indicates that some weakly correlated connections are eliminated using ReLU matrix decomposition, (·) T represents the transpose. As an example, T w1 and T w2 They represent the embeddings of the source node and the target node in the graph structure representation 502 respectively.

[0085] In some embodiments, the feature representation extraction module 501 may use the adaptive attention matrix The transfer matrix A regarded as an implicit diffusion process att , the output D of the graph attention layer at the i-th time slice in the graph diffusion convolution att (X (i) ,A att ) can be expressed by formula (6).

[0086]

[0087] Among them, D O =diag(A att ), D I =diag(A att T ), and They represent the transfer matrix of the bidirectional diffusion process, W k1 and W k2 is a learnable parameter, W k1 ∈R 1×dk , W k2 ∈R 1×dk , dk represents the first embedding dimension, and diag(·) represents the diagonal operation. Through the graph attention layer D att (X (i) ,A att ) After extracting the spatial dependency relationship between the multiple traffic features 509, the graph structure representation 502 corresponding to each historical time slice can be obtained. The graph structure representation 502 can also be represented as H i , H i ∈R N×dk In essence, the graph structure representation 502 implicitly learns the spatial dependencies between multiple traffic features 509 by aggregating each hop neighbor nodes on the graph.

[0088] In addition to the spatial dependencies between multiple traffic features 509, each traffic feature sequence 508 also contains a time dependency. This time dependency not only shows the contextual relationship, but also has a global impact. In an embodiment of the present disclosure, the feature representation extraction module 501 can extract the time dependency (or also referred to as global time dependency) between multiple traffic features 509 based on a first machine learning model 503 with a multi-head attention mechanism. As an example, the first machine learning model 503 can be any appropriate model, including but not limited to a Transformer model, etc.

[0089] As an example, the feature representation extraction module 501 can connect the graph structure representation H corresponding to all historical time slices i, thus obtaining the graph structure representation H f ∈R B×N×T×dk , where B is the batch size. Next, the feature representation extraction module 501 extracts the graph structure representation H f The first two dimensions of are flattened to obtain the graph structure representation E f ∈R BN×T×dk Then, the feature representation extraction module 501 uses multiple (for example, three) transformation matrices to represent the flattened graph structure E f Perform linear projection to obtain matrices Q, K, and V, which represent query features, key features, and value features, respectively.

[0090] As an example, the output A of the self-attention layer of the first machine learning model 503 att (Q, K, V) can be expressed by formula (7).

[0091]

[0092] Among them, Q, K, V ∈ R BN×T×dw , dw represents the second embedding dimension. In order to enhance the expression capability, the first machine learning model 503 is provided with a multi-head attention mechanism to extract the temporal dependency between multiple traffic features 509 in multiple subspaces.

[0093] As an example, the output MultiAtt(Q, K, V) of the multi-head attention mechanism can be expressed by Formula (8) and Formula (9).

[0094] MultiAtt(Q,K,V)=Concat(head 1 , head 1 , ..., head 1 ); (8)

[0095] head 1 =A att (E f Θ Q , E f Θ K , E f Θ V ); (9)

[0096] Where n is the number of attention heads, Θ Q , Θ K , Θ V is the projection matrix, Θ Q , Θ K , Θ V ∈R dk×dw .

[0097] Next, the feature representation extraction module 501 can input the output of the multi-head attention mechanism MultiAtt(Q, K, V) into the feed-forward network with residual connection to obtain the first feature representation E seq , E seq ∈R B ×NT×demb , demb represents the third embedding dimension. In this way, the embodiment of the present disclosure learns the hidden spatiotemporal dependencies between multiple traffic features 509 through graph diffusion convolution and multi-head attention mechanism.

[0098] It should be noted that the above algorithms and parameters are only exemplary and do not constitute a limitation on the embodiments of the present disclosure. According to actual needs, the embodiments of the present disclosure can adopt any appropriate algorithm and parameters, which will not be listed here one by one.

[0099] In some embodiments, the content recommendation system 150 may use the second machine learning model 302 to determine the predicted traffic changes of the live stream 403 corresponding to each of the multiple interactive operations 402 in the target time slice based on multiple traffic feature sequences 508 and sparse feature information 5011 related to the live stream 403. Alternatively or additionally, the content recommendation system 150 may use the second machine learning model 302 to determine the predicted traffic changes of the live stream 403 corresponding to each of the multiple interactive operations 402 in the target time slice based on multiple traffic feature sequences 508 and dense feature information 5012 related to the live stream 403 and / or multiple interactive operations 402. In other words, the input of the second machine learning model 302 may include the above-mentioned sparse feature information 5011 and / or dense feature information 5012 in addition to the multiple traffic feature sequences 508.

[0100] Continue to refer to Figure 5 , Figure 5 A plurality of predicted flow rate changes 5016-1, 5016-2, 5016-j are shown, where j is a positive integer. For ease of discussion, the plurality of predicted flow rate changes 5016-1, 5016-2, 5016-j are collectively or individually referred to as predicted flow rate changes 5016 below. As an example, Figure 5 The multiple predicted traffic changes 5016 - 1 , 5016 - 2 , and 5016 - j shown in the figure may be predicted traffic changes 5016 corresponding to comment operations, like operations, and sending virtual gifts, respectively.

[0101] As an example, sparse features may refer to features that do not appear continuously in a data set, and most of the values ​​are zero. As an example, the sparse feature information 5011 may include the basic live broadcast room features 5011-1 of the live broadcast room to which the live broadcast stream 403 belongs, the basic anchor features 5011-2 of the anchor to which the live broadcast stream 403 belongs, the anchor statistical features 5011-3 associated with the anchor in the live broadcast stream 403, and the live broadcast room statistical features 5011-4 associated with the live broadcast room, etc. As an example, the basic live broadcast room features include but are not limited to the live broadcast room type, live broadcast room tags, live broadcast area, etc. The anchor basic features may include anchor type, anchor tags, etc. The live broadcast room statistical features 5011-4 may include the cumulative number and / or average number of one or more interactive operations 402 initiated by the broadcast viewing user 401 in a predetermined number of time slices, etc. The predetermined number here can be set according to actual needs, for example, the predetermined number can be set to 1 / 3 / 5 / 10 / 15, etc. The live broadcast room statistical feature 5011-4 may also include the cumulative number and / or average number of one or more interactive operations 402 initiated by the viewing user 401 in the live broadcast room in the past predetermined time period. Alternatively or additionally, the anchor statistical feature 5011-3 may also include the anchor's average live broadcast duration and / or live broadcast days in the past predetermined time period. The predetermined time period here can be set according to actual needs, for example, the predetermined time period can be set to 1 / 7 / 14 days, etc.

[0102] As an example, dense features may refer to features that frequently appear in a data set. In an embodiment of the present disclosure, the dense feature information 5012 may include, for example, more fine-grained statistical features, etc. For example, the dense feature information 5012 includes statistical features for the traffic feature sequence 508 (e.g., the number of likes, the like rate, the number of comments, etc. within the last several time slices). In addition, the dense feature information 5012 may also include statistical features such as the cumulative number and / or average number of interactive operations 402 collected using a finer time window.

[0103] As an example, the sparse feature information 5011 may further include multimodal features 5011-5. The multimodal features 5011-5 may include, for example, image features and text features, etc. The multimodal features 5011-5 may be obtained by means such as automatic speech recognition (ASR) and / or optical character recognition (OCR). As an example, the content recommendation system 150 may encode the results of automatic speech recognition and / or the results of optical character recognition to generate an image feature representation and a text feature representation. In some embodiments, the content recommendation system 150 may align the image feature representation and the text feature representation of the same live broadcast room based on a contrast learning method for segments, so as to enhance the multimodal representation effect. The content recommendation system 150 may also refine the representations of different live streams 403 that the same viewing user 401 is interested in, etc. After obtaining the image feature representation and the text feature representation, the content recommendation system 150 may use a deep clustering algorithm to generate hierarchical multimodal features 5011-5. As an example, the multimodal features 5011-5 may be updated based on a predetermined period. The update based on the predetermined period may be determined based on the length of the time slice to capture the real-time content changes of the live stream 403.

[0104] As an example, the content recommendation system 150 may convert the identification feature 5013 (such as the host identifier of the host to which the live stream 403 belongs and the slice identifier of the time slice to which the sparse feature information 5011 belongs) and the sparse feature information 5011 into a unified low-dimensional dense feature representation through the feature encoding unit 5014, and then splice the converted low-dimensional dense feature representation. Thereby, it is convenient for the second machine learning model 302 to learn the personalized interaction traffic change patterns of different time slices (and / or different hosts).

[0105] As an example, the content recommendation system 150 may use the dense feature processing module 5017 to process the dense feature information 5012 to obtain a better feature representation. For example, the dense feature processing module 5017 may include, but is not limited to, deep neural networks (DNN), etc. Subsequently, the content recommendation system 150 may use the connection module 5015 to splice the output of the feature encoding unit 5014, the output of the dense feature processing module 5017, and the output of the feature representation extraction module 501 to obtain the splicing result E v , the splicing result E v can be expressed as E v =[E basic ,E stat ,E seq ,E mm ,E id , where E basic ,E stat ,E seq,E mm ,E id They respectively represent dense feature representations determined based on basic features (such as live broadcast room basic features 5011-1 and anchor basic features 5011-2, etc.), statistical features (such as live broadcast room statistical features 5011-4, anchor statistical features 5011-3 and statistical features involved in dense feature information 5012, etc.), traffic feature sequence 508 and identification features 5013.

[0106] Then, the content recommendation system 150 can perform a cross-network analysis on the concatenated result E v Perform high-order feature crossover, and then use the second machine learning model 302 (such as a multi-task learning model) based on the feature sharing mechanism to determine the predicted traffic changes 5016 corresponding to each of the multiple interactive operations 402. As an example, the content recommendation system 150 can implement feature crossover based on a deep neural network (DNN). In addition, in some embodiments, the feature sharing component in the multi-task learning model can also be replaced by other networks (such as DCN-V2 or XDeepFM, etc.). As an example, the main task of the second machine learning model 302 is to maximize the likelihood probability between the predicted value and the actual value. Therefore, the embodiments of the present disclosure can be achieved through the standard cross entropy loss function To train the second machine learning model 302. As an example, the loss function It can be expressed by formula (10).

[0107]

[0108] Where Θ represents the learnable parameters of the second machine learning model 302, and denote respectively the predicted traffic change 5016 and the actual traffic change of the j-th interactive operation 402 at the i-th time slice, represents the L2 regularization term, and γ represents the hyperparameter used for balancing.

[0109] In this way, the embodiment of the present disclosure utilizes the feature representation extraction module 501 to perform feature encoding on the traffic feature sequence 508. Then, after converting all sparse feature representations into dense feature representations, DNN is used to perform high-order crossover on all dense feature representations. Finally, the embodiment of the present disclosure utilizes a multi-task learning model with feature sharing capability to obtain predicted traffic changes 5016 corresponding to each of the multiple interactive operations 402. Thus, the embodiment of the present disclosure can extract potential spatiotemporal dependencies between multiple traffic features 509. In addition, the embedding of the anchor identifier and the time slice identifier enables the second machine learning model 302 to simultaneously learn the personalized information of the anchor and the time slice.

[0110] It should be noted that the above algorithms and parameters are only exemplary and do not constitute a limitation on the embodiments of the present disclosure. According to actual needs, the embodiments of the present disclosure can adopt any appropriate algorithm and parameters, which will not be listed here one by one.

[0111] Once the predicted traffic changes 5016 of the live stream 403 corresponding to each of the multiple interactive operations 402 in the target time slice are determined, in box 230, the content recommendation system 150 determines the recommendation level of the live stream 403 for the user group based at least on these predicted traffic changes 5016.

[0112] As an example, the content recommendation system 150 can divide the user groups based on the viewing habits of the viewing users 401. For example, the content recommendation system 150 can divide the viewing users 401 with similar interests into the same user group. From the perspective of the viewing users 401, the viewing users 401 are concerned about the attractiveness of the live content. Therefore, the content recommendation system 150 can determine the recommendation degree of the live stream 403 for the user group based on the viewing habits and other characteristics of the user group and the predicted traffic change 5016. As an example, the content recommendation system 150 can determine one or more candidate live streams for the user group based on the viewing habits and other characteristics of the user group. Then, the content recommendation system 150 further adjusts the recommendation degree of these live streams 403 based on the predicted traffic change 5016 of each candidate live stream to change the recommendation priority of these live streams 403. For example, the higher the predicted traffic, the more popular the live stream 403 will be in the future, so the recommendation degree of the live stream 403 can be increased accordingly.

[0113] In some embodiments, the content recommendation system 150 may determine an adjusted recommendation score for the live stream 403 based on the predicted traffic changes 5016 corresponding to each of the multiple interactive operations 402, wherein the adjusted recommendation score indicates the overall traffic change of the live stream 403 corresponding to the multiple interactive operations 402 in the target time slice. Then, the content recommendation system 150 may determine the recommendation degree of the live stream 403 for the user group (also known as the viewing user 401) based on the reference recommendation score of the live stream 403 for the user group and the adjusted recommendation score.

[0114] As an example, the adjustment recommendation score may be calculated by any appropriate method such as weighted summing of the predicted traffic changes 5016, in order to ensure that the impact of traffic changes of different interactive operations 402 on the final score is reasonable. The reference recommendation score may be determined based on factors such as the content quality of the live stream 403, the popularity of the anchor, and the historical viewing records of the viewing user 401. As an example, the recommendation degree here may be a specific value or ranking, and the recommendation degree is used to indicate the recommendation priority of the live stream 403 in the user group.

[0115] By adjusting the recommendation score and the degree of recommendation, the content recommendation system 150 can ensure that the anchor's live stream 403 obtains sufficient exposure and interaction within the appropriate time slice. This helps to increase the anchor's popularity and revenue. At the same time, the content recommendation system 150 will also consider the viewing experience and satisfaction of the viewing user 401. By recommending the live stream 403 that the viewing user 401 is interested in, the content recommendation system 150 can increase the user's viewing time and interaction frequency. Therefore, the embodiments of the present disclosure achieve a virtuous cycle of the live broadcast ecosystem by comprehensively considering the bilateral experience of the anchor and the viewing user 401.

[0116] In some embodiments, the degree of recommendation is positively correlated with the overall traffic change indicated by the adjustment of the recommendation score.

[0117] As an example, the recommendation level R exp It can be expressed by formula (11) and formula (12).

[0118] R exp＝ R base +G score ; (11)

[0119]

[0120] Among them, R base represents the reference recommendation score for the user group, G score represents the adjusted recommendation score, and α and β are hyperparameters for combining all predicted traffic changes 5016. In this way, the embodiment of the present disclosure models the trend of interactive traffic changes (or traffic efficiency) from the perspective of the live broadcast room and takes time slices as the granularity, thereby providing an additional information gain factor for the recommendation of the live broadcast stream 403.

[0121] During the live broadcast process, the anchor is mainly concerned about the stability of the interactive traffic, the total traffic number and the traffic performance. The content recommendation system 150 can allocate a certain amount of effective push traffic to the live stream 403 within a unit time slice based on the online service through the proportional-integral-differential control system (PID). Effective push traffic refers to push traffic that can bring actual interaction and conversion value to the live stream 403. For the anchor, effective push traffic means more audience participation and higher exposure. However, effective traffic resources are scarce, and the demand for effective push traffic is different at different stages of the live broadcast process. In order to ensure that the anchor with improved live broadcast quality can obtain certain effective push traffic incentives in a timely manner, the embodiments of the present disclosure can further combine the predicted traffic change 5016 to adjust the allocation of effective push traffic.

[0122] In some embodiments, the content recommendation system 150 may adjust the target push traffic associated with the live stream 403 in the target time slice based on the predicted traffic changes 5016 corresponding to each of the multiple interactive operations 402. Then, the content recommendation system 150 may determine the traffic control score of the live stream 403 in the target time slice based at least on the adjusted target push traffic and the consumed push traffic associated with the live stream 403. Subsequently, the content recommendation system 150 may determine the ranking of the live stream 403 in the live push sequence of the user group based at least on the traffic control score and the recommendation degree of the live stream 403.

[0123] As an example, the target push traffic may refer to the effective push traffic that the content recommendation system 150 plans to allocate to the live stream 403 within the target time period. The consumed push traffic may refer to the effective push traffic that the content recommendation system 150 has provided to the live stream 403 as of the current moment. As an example, the target push traffic and the consumed push traffic may be represented in the form of counts. For example, assuming that the content recommendation system 150 plans to provide "2" likes for the live stream 403, then the target push traffic may be recorded as "2". Assuming that there is currently a user accessing the live stream 403 based on the effective push traffic and initiating a like operation in the live stream 403, then the consumed push traffic may be recorded as "1". It should be noted that the above description of the counting of the target push traffic and the consumed push traffic is only exemplary content, which does not constitute a limitation on the embodiments of the present disclosure. According to actual needs, the target push traffic and the consumed push traffic may also be counted in other ways.

[0124] As an example, when the recommendation degree remains unchanged, the traffic control score is positively correlated with the ranking of the live broadcast stream 403 in the live broadcast push sequence of the user group. That is, the larger the traffic control score, the higher the ranking of the live broadcast stream 403 in the live broadcast push sequence of the user group, and the smaller the traffic control score, the lower the ranking of the live broadcast stream 403 in the live broadcast push sequence of the user group.

[0125] As an example, the traffic control score is related to the difference between the target push traffic and the consumed push traffic. Specifically, when the consumed push traffic remains unchanged, the target push traffic is positively correlated with the traffic control score, that is, the larger the target push traffic, the larger the traffic control score, and the smaller the target push traffic, the smaller the traffic control score. In an embodiment of the present disclosure, when the predicted traffic change 5016 indicates that the interactive traffic of the live stream 403 is about to increase, the target push traffic can be increased (for example, from "2" to "3") to improve the traffic control score, thereby appropriately improving the ranking of the live stream 403 in the live push sequence of the user group. In other cases, the target push traffic can be gradually approached to the initial value before adjustment (for example, from "3" to "2"), so that the traffic control score of the live stream 403 whose interactive traffic is stabilizing or about to decrease is restored to the baseline level.

[0126] Figure 6 FIG. 6 is a schematic diagram showing an example 600 of adjusting a recommendation ranking according to some embodiments of the present disclosure. Figure 6 , the content recommendation system 150 also includes a decision module 601. The decision module 601 is used to adjust the traffic control score to enhance the live broadcast experience of the anchor. The decision module 601 can trigger related services (such as statistical services 602) in real time or periodically to obtain (6101) the consumed push traffic (also referred to as the consumed push traffic count) associated with the live stream 403. Then, the decision module 601 obtains an error signal by subtracting the consumed push traffic count from the set target push traffic (also referred to as the target push traffic count). Then, based on the error signal, the traffic control score is determined.

[0127] As an example, the flow control score μ p The adjustment process can be expressed by formula (13) to formula (15).

[0128]

[0129]

[0130] μ p =intergal+lagboost; (15)

[0131] Where N target -N impr Represents the error signal, N target Indicates the target push traffic, N impr Indicates the consumed push traffic, indicates the current historical time slice, α I , α decay , t targetThey are respectively the first predetermined coefficient, the second predetermined coefficient and the third predetermined coefficient related to the flow control score regulation, and n represents the window size, that is, n historical time slices.

[0132] In some embodiments, the content recommendation system 150 may determine an adjusted recommendation score for the live stream 403 based on the predicted traffic changes 5016 corresponding to each of the multiple interactive operations 402, wherein the adjusted recommendation score indicates the overall traffic change of the live stream 403 corresponding to the multiple interactive operations 402 in the target time slice. Then, in response to the adjusted recommendation score indicating that the overall traffic of the live stream 403 corresponding to the multiple interactive operations 402 in the target time slice is about to increase, the content recommendation system 150 may increase the target push traffic by a first value. Alternatively or additionally, in response to the adjusted recommendation score indicating that the overall traffic of the live stream 403 corresponding to the multiple interactive operations 402 in the target time slice is about to decrease, the content recommendation system 150 may reduce the target push traffic by a second value.

[0133] As an example, the content recommendation system 150 can dynamically adjust the target push traffic based on the predicted traffic change 5016. This design is intended to enable the content recommendation system 150 to enable the live stream 403 to obtain more effective push traffic in a timely manner when the predicted traffic of the live stream 403 is about to increase. When the predicted traffic of the live stream 403 is about to decrease, the content recommendation system 150 reduces the effective push traffic. When the predicted traffic of the live stream 403 is in a stable state, the content recommendation system 150 ensures that the target push traffic Ntarget can quickly converge to the baseline level. In this way, the dynamic allocation of effective push traffic in the spatiotemporal domain can be achieved.

[0134] The following is a description of the process of adjusting the target push traffic in the embodiment of the present disclosure.

[0135] Continue with formula (13) to formula 15 in the previous text. First, initialize the following parameters related to the adjustment process of the target push traffic: R - =0, R + =0, S dura =∞,S st =0,N + =0,N - =0, where R - Indicates that the predicted flow change 5016 of time slice t will decrease, R + Indicates that the predicted flow change 5016 of time slice t will increase, S dura Indicates the duration of the predicted flow increase / decrease state, S st Indicates the predicted start time of the flow increase / decrease state, N + Represents R + The count, N - Represents R- It should be noted that R + , R - is unique within each time slice.

[0136] Then, let t be 1, 2, ..., T in sequence, and calculate the value of the target push flow each time t takes a value, where T is the total number of time slices. As an example, the calculation process of the target push flow each time t takes a value can be expressed based on formulas (16) to (23).

[0137] S dura =tS st ; (16)

[0138] If S dura > gap & G score ≥g + , then R + =1, S st =t,N + +1; (17)

[0139] If S dura > gap & G score ≤g - , then R - =1, S st =t,N - +1; (18)

[0140] If S dura ≤T gap or g - <G score <g + , then R + =R - =0, S st =0; (19)

[0141] Δmpc=R + -R - ; (20)

[0142] δt=(tS st ); (twenty one)

[0143]

[0144] N target =N target +Γ impr ×Δmpc×Τ decay ; (twenty three)

[0145] where Γimpr ×Δmpc×Τ decay is the first value of the increased target push traffic or the second value of the decreased target push traffic, Τ gap , Τ decay , Γ impr represents the hyperparameter for error signal correction, and g+ and g- respectively represent the upper threshold and the lower threshold for adjusting the recommendation score G score . g+ and g- can be set according to the precision-recall threshold.

[0146] Once the adjusted target push traffic and the consumed push traffic are determined, the decision module 601 can calculate the outputs of the proportional term and the integral term through the PID system. Subsequently, the decision module 601 sends the output of the PID system (such as the traffic control score) to the live forward indexing service 603. The client devices 604 of the user group can send recommendation requests 605 to the content recommendation system 150 in real time or periodically. The live forward indexing service 603 in the content recommendation system 150 can determine the live stream 403 to be recommended to the user group through multiple sorting stages. As an example, the live forward indexing service 603 can determine the live stream 403 (i.e., the live push sequence 608) to be recommended to the user group through at least a first sorting stage 606 (also referred to as the rough sorting stage) and a second sorting stage 607 (also referred to as the fine sorting stage) performed based on the sorting result of the first sorting stage 606. As an example, in each sorting stage, the live forward indexing service 603 can determine the ranking of the live stream 403 in this sorting stage based on the recommendation degree of the live stream 403 (determined based on multiple predicted traffic changes output by the second machine learning model) and / or the traffic control score output by the decision module 601.

[0147] In this way, the embodiments of the present disclosure can take into account the bilateral experiences of the viewing users 401 and the live streamers. Ensure that when the live performance of the live streamer improves, high-quality effective traffic incentives can be obtained. At the same time, the viewing users 401 can watch more high-quality live content, and through the joint optimization of the bilateral experiences, the live recommendation effect is improved.

[0148] In some embodiments, the trigger service 609 in the content recommendation system 150 can trigger the feature service 305 to generate traffic features based on a third predetermined period 6011 (such as 10 seconds or other appropriate time), and determine the predicted traffic change 5016 by means of the second machine learning model 302. Alternatively or additionally, the trigger service 609 can also trigger the decision module 601 based on the third predetermined period 6011, so as to determine the traffic control score.

[0149] In some embodiments, after the viewing user 401 accesses the live stream 403 based on the live push sequence 608, the traffic statistics module 6010 can update the count of the consumed push traffic in the statistics unit 602 and the current interactive traffic of the live stream 403 based on the interactive operations taken by the viewing user 401 in the live stream 403. In this way, the content recommendation system 150 can continuously collect the real-time interactive traffic changes in the live stream 403, so as to update the learnable parameters of the second machine learning model 302 in real time, and ensure that the second machine learning model 302 can adapt to the new traffic distribution.

[0150] According to the various embodiments described above, it can be clearly understood that the embodiments of the present disclosure construct a data stream processing module 301 (also referred to as a data stream engine), which can support the attribution of interactive traffic changes at the granularity of time slices from the perspective of the live broadcast room. At the same time, the embodiments of the present disclosure also propose a feature extraction module 501 (also referred to as a spatiotemporal fusion module) to capture the dynamic trend of interactive traffic changes. The feature extraction module 501 uses rich basic features, statistical features, traffic feature sequences 508, etc. to mine the potential spatiotemporal dependencies between multiple interactive operations 402. Finally, the embodiments of the present disclosure use a decision module 601 to integrate the traffic prediction results into the online recommendation process of the live broadcast stream 403, thereby balancing the benefits of the consumption side and the supply side from the bilateral perspectives of the anchor and the viewing user 401.

[0151] The embodiments of the present disclosure also provide corresponding devices for implementing the above methods or processes. Figure 7 A schematic structural block diagram of an apparatus 700 for live broadcast recommendation according to some embodiments of the present disclosure is shown. The apparatus 700 may be implemented as or included in a content recommendation system 150. Each module / component in the apparatus 700 may be implemented by hardware, software, firmware, or any combination thereof.

[0152] Reference Figure 7The device 700 includes a traffic feature sequence determination module 710, a predicted traffic change determination module 720, and a recommendation degree determination module 730. The traffic feature sequence determination module 710 is configured to determine a plurality of traffic feature sequences corresponding to a plurality of interactive operations of a live stream, wherein the plurality of interactive operations are initiated by viewing users in the live stream, and in each of the plurality of traffic feature sequences, different traffic features represent feature information corresponding to the interactive operations in different historical time slices of the live stream. The predicted traffic change determination module 720 is configured to determine, based at least on the plurality of traffic feature sequences, predicted traffic changes corresponding to each of the plurality of interactive operations of the live stream in a target time slice. The recommendation degree determination module 730 is configured to determine the recommendation degree of the live stream for the user group based at least on the recommendation request of the user group and the predicted traffic changes corresponding to each of the plurality of interactive operations of the live stream in the target time slice.

[0153] In some embodiments, the predicted traffic change determination module 720 is further configured to: determine the dependency relationship between each traffic feature in a plurality of traffic feature sequences and other traffic features other than the traffic feature; extract a first feature representation of the plurality of traffic feature sequences based on the determined dependency relationship; and determine the predicted traffic changes corresponding to each of the plurality of interactive operations in the target time slice of the live stream based at least on the first feature representation.

[0154] In some embodiments, the dependency between at least one traffic feature and other traffic features other than the traffic feature includes at least a first dependency and a second dependency; and the first dependency indicates the dependency between traffic features corresponding to different interactive operations in the same historical time slice, and the second dependency indicates the dependency between traffic features belonging to different historical time slices in the same traffic feature sequence.

[0155] In some embodiments, the predicted traffic change determination module 720 is further configured to: based on the determined dependency, extract multiple graph structure representations corresponding to multiple historical time slices of the live stream from multiple traffic feature sequences, wherein for a given graph structure representation corresponding to a given historical time slice, the nodes in the given graph structure representation correspond to the traffic features belonging to the given historical time slice in the multiple traffic feature sequences, and the edge between at least two nodes in the given graph structure representation is determined based on the dependency between the corresponding two traffic features; and determine a first feature representation through feature encoding based on the dependency between the multiple graph structure representations.

[0156] In some embodiments, multiple graph structure representations are extracted from multiple traffic feature sequences using at least graph diffusion convolution; and / or the dependencies between the multiple graph structure representations are determined using at least a first machine learning model with a multi-head attention mechanism.

[0157] In some embodiments, the predicted traffic change determination module 720 is further configured to: determine the predicted traffic changes corresponding to each of the multiple interactive operations in the live stream in the target time slice using a second machine learning model based on multiple traffic feature sequences and at least one of the following: discrete feature information related to the live stream, or dense feature information related to the live stream and / or multiple interactive operations.

[0158] In some embodiments, the recommendation degree determination module 730 is further configured to: determine an adjusted recommendation score for the live stream based on the predicted traffic changes corresponding to each of the multiple interactive operations, wherein the adjusted recommendation score indicates the overall traffic change of the live stream corresponding to the multiple interactive operations in the target time slice; and determine the recommendation degree of the live stream for the user group based on the reference recommendation score of the live stream for the user group and the adjusted recommendation score.

[0159] In some embodiments, the degree of recommendation is positively correlated with the overall traffic change indicated by the adjustment of the recommendation score.

[0160] In some embodiments, the apparatus 600 further includes a ranking control module. The ranking control module is configured to: adjust the target push traffic associated with the live stream in the target time slice based on the predicted traffic changes corresponding to each of the multiple interactive operations; determine the traffic control score of the live stream in the target time slice based at least on the adjusted target push traffic and the consumed push traffic associated with the live stream; and determine the ranking of the live stream in the live push sequence of the user group based at least on the traffic control score and the recommendation degree of the live stream.

[0161] In some embodiments, the ranking control module is configured to: determine an adjusted recommendation score for the live stream based on predicted traffic changes corresponding to each of the multiple interactive operations, wherein the adjusted recommendation score indicates the overall traffic change of the live stream corresponding to the multiple interactive operations in the target time slice; increase the target push traffic by a first value in response to the adjusted recommendation score indicating that the overall traffic corresponding to the multiple interactive operations of the live stream in the target time slice is about to increase; and reduce the target push traffic by a second value in response to the adjusted recommendation score indicating that the overall traffic corresponding to the multiple interactive operations of the live stream in the target time slice is about to decrease.

[0162] Figure 8 8 is a block diagram of an electronic device 800 in which one or more embodiments of the present disclosure may be implemented. The electronic device 800 may be used to implement, for example, Figure 1The content recommendation system 150 shown or Figure 7 The device 700 shown. It should be understood that Figure 8 The electronic device 800 shown is merely exemplary and should not constitute any limitation on the functionality and scope of the embodiments described herein.

[0163] Reference Figure 8 , the electronic device 800 is in the form of a general electronic device. The components of the electronic device 800 may include, but are not limited to, one or more processors or processors 810, a memory 820, a storage device 830, one or more communication units 840, one or more input devices 850, and one or more output devices 860. The processor 810 may be an actual or virtual processor and is capable of performing various processes according to a program stored in the memory 820. In a multi-processor system, multiple processors execute computer executable instructions in parallel to improve the parallel processing capability of the electronic device 800.

[0164] The electronic device 800 typically includes a plurality of computer storage media. Such media can be any available media accessible to the electronic device 800, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 820 can be a volatile memory (e.g., register, cache, random access memory (RAM)), a non-volatile memory (e.g., a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), flash memory) or some combination thereof. The storage device 830 can be a removable or non-removable medium, and can include a machine-readable medium, such as a flash drive, a disk, or any other medium, which can be used to store information and / or data and can be accessed within the electronic device 800.

[0165] The electronic device 800 may further include additional removable / non-removable, volatile / non-volatile storage media. Figure 8 As shown in , a disk drive for reading or writing from a removable, non-volatile disk (e.g., a "floppy disk") and an optical drive for reading or writing from a removable, non-volatile optical disk may be provided. In these cases, each drive may be connected to the bus (not shown) by one or more data media interfaces. The memory 820 may include a computer program product 825 having one or more program modules that are configured to perform various methods or actions of various embodiments of the present disclosure.

[0166] The communication unit 840 implements communication with other electronic devices through a communication medium. Additionally, the functions of the components of the electronic device 800 can be implemented with a single computing cluster or multiple computing machines that can communicate through a communication connection. Therefore, the electronic device 800 can operate in a networked environment using a logical connection with one or more other servers, a network personal computer (PC), or another network node.

[0167] The input device 850 may be one or more input devices, such as a mouse, a keyboard, a tracking ball, etc. The output device 860 may be one or more output devices, such as a display, a speaker, a printer, etc. The electronic device 800 may also communicate with one or more external devices (not shown) through the communication unit 840 as needed, such as a storage device, a display device, etc., communicate with one or more devices that allow a user to interact with the electronic device 800, or communicate with any device that allows the electronic device 800 to communicate with one or more other electronic devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0168] According to an exemplary implementation of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method described above.

[0169] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices, equipment, and computer program products implemented according to the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0170] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so as to produce a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device for implementing the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions make the computer, programmable data processing device, and / or other equipment work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0171] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0172] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple implementations of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some implementations as replacements, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions.

[0173] The above descriptions of various implementations of the present disclosure are exemplary, non-exhaustive, and not limited to the disclosed implementations. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The determination of the terms used herein is intended to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the various implementations disclosed herein.

Claims

1. A live broadcast recommendation method, comprising: Determine a plurality of traffic feature sequences corresponding to a plurality of interactive operations of the live stream, wherein the plurality of interactive operations are initiated by a viewing user in the live stream, and in each of the plurality of traffic feature sequences, different traffic features represent feature information corresponding to the interactive operations in different historical time slices of the live stream; Determine, based at least on the multiple traffic feature sequences, predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in a target time slice; as well as The recommendation degree of the live stream for the user group is determined based at least on the recommendation request of the user group and the predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in the target time slice.

2. The method according to claim 1, wherein determining the predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in the target time slice based at least on the multiple traffic feature sequences comprises: Determine a dependency relationship between each flow feature in the plurality of flow feature sequences and other flow features except the flow feature; Based on the determined dependency relationship, extracting a first feature representation of the plurality of traffic feature sequences; as well as Based at least on the first feature representation, the predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in the target time slice are determined.

3. The method according to claim 2, wherein the dependency relationship between the at least one flow feature and other flow features other than the at least one flow feature comprises at least a first dependency relationship and a second dependency relationship; and The first dependency indicates the dependency between traffic features corresponding to different interactive operations in the same historical time slice, and the second dependency indicates the dependency between traffic features belonging to different historical time slices in the same traffic feature sequence.

4. The method according to claim 2, wherein extracting the first feature representation of the plurality of traffic feature sequences comprises: Based on the determined dependency relationship, extracting multiple graph structure representations corresponding to multiple historical time slices of the live stream from the multiple traffic feature sequences, wherein for a given graph structure representation corresponding to a given historical time slice, a node in the given graph structure representation corresponds to a traffic feature belonging to the given historical time slice in the multiple traffic feature sequences, and an edge between at least two nodes in the given graph structure representation is determined based on the dependency relationship between the corresponding two traffic features; as well as Based on the dependency relationship between the multiple graph structure representations, the first feature representation is determined through feature encoding.

5. The method according to claim 4, wherein the plurality of graph structure representations are extracted from the plurality of traffic feature sequences using at least graph diffusion convolution; and / or The dependency relationship between the multiple graph structure representations is determined by at least using a first machine learning model with a multi-head attention mechanism.

6. The method according to claim 1, wherein determining the predicted traffic changes of the live stream corresponding to each of the plurality of interactive operations in the target time slice comprises: Based on the multiple traffic feature sequences and at least one of the following, a second machine learning model is used to determine predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in the target time slice: Discrete characteristic information related to the live stream, or Dense feature information related to the live stream and / or the plurality of interactive operations.

7. The method according to claim 1, wherein determining the recommendation level of the live stream to the user group comprises: Determine, based on the predicted traffic changes corresponding to each of the multiple interactive operations, a recommended score for adjusting the live stream, wherein the recommended score for adjusting indicates an overall traffic change of the live stream corresponding to the multiple interactive operations in the target time slice; as well as The recommendation degree of the live stream for the user group is determined based on the reference recommendation score of the live stream for the user group and the adjusted recommendation score. The method according to claim 7 , wherein the recommendation level is positively correlated with the overall flow change indicated by the adjusted recommendation score.

9. The method according to claim 1, further comprising: Adjusting a target push traffic associated with the live stream in the target time slice based on predicted traffic changes corresponding to each of the plurality of interaction operations; Determining a flow control score of the live stream in the target time slice based at least on the adjusted target push flow and the consumed push flow associated with the live stream; as well as Based at least on the traffic control score of the live stream and the recommendation degree, a ranking of the live stream in the live broadcast push sequence of the user group is determined.

10. The method according to claim 9, wherein adjusting the target push traffic associated with the live stream comprises: Determine, based on the predicted traffic changes corresponding to each of the multiple interactive operations, a recommended score for adjusting the live stream, wherein the recommended score for adjusting indicates an overall traffic change of the live stream corresponding to the multiple interactive operations in the target time slice; In response to the adjusted recommendation score indicating that the overall traffic of the live stream corresponding to the multiple interactive operations in the target time slice is about to increase, increasing the target push traffic by a first value; as well as In response to the adjusted recommendation score indicating that the overall traffic of the live stream corresponding to the multiple interactive operations in the target time slice is to be reduced, the target push traffic is reduced by a second value.

11. A device for live broadcast recommendation, comprising: A traffic feature sequence determination module is configured to determine a plurality of traffic feature sequences corresponding to a plurality of interactive operations of a live stream, wherein the plurality of interactive operations are initiated by a viewing user in the live stream, and in each of the plurality of traffic feature sequences, different traffic features represent feature information corresponding to the interactive operations in different historical time slices of the live stream; A predicted traffic change determination module is configured to determine predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in a target time slice based at least on the multiple traffic feature sequences; as well as The recommendation degree determination module is configured to determine the recommendation degree of the live stream for the user group based at least on the recommendation request of the user group and the predicted traffic changes of the live stream corresponding to each of the multiple interactive operations in the target time slice.

12. An electronic device comprising: at least one processor; as well as At least one memory, the at least one memory is coupled to the at least one processor and stores instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processor, causing the electronic device to perform the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having computer-executable instructions stored thereon, wherein the computer-executable instructions can be executed by a processor to implement the method according to any one of claims 1 to 10.

14. A computer program product comprising computer executable instructions, wherein the computer executable instructions, when executed by a processor, implement the method according to any one of claims 1 to 10.

Images