JP7052768B2 - Distribution design support method, distribution design support device and program - Google Patents

Description

The present invention relates to a distribution design support method, a distribution design support device, and a program.

With the development of codec technology, video distribution technology, and display technology, adaptive bit rate video distribution services for mobile terminals have rapidly become widespread. The adaptive bit rate video distribution service enables smooth video playback by distributing video data at the optimum bit rate according to the communication status even in an environment where the network status fluctuates from moment to moment, such as a mobile environment.

On the other hand, adaptive bit rate video distribution does not always guarantee smooth video playback, and the encoding quality deteriorates and playback stops due to an extreme decrease in throughput and exhaustion of the buffer of the receiving terminal. , User experience quality (QoE: Quality of Experience) may be significantly reduced. Quality deterioration not only reduces QoE, but also reduces viewing time and cancels services. Therefore, it is necessary for video distribution service providers to quantitatively evaluate these indicators and manage them to maintain a certain level. Will be.

Now, suppose that viewing time is assumed as a management target index. Factors that influence the viewing time include content factors (type, content length, ranking, etc.) corresponding to each of the content, distribution server, network, and terminal, which are components of the video distribution service, and coding performance factors (bit rate, frame). Service factors such as rate, resolution, etc.), network performance factors (throughput, packet delay fluctuation, packet loss, packet delay, etc.), terminal performance factors (buffer size, etc.) and user factors such as gender and age can be considered. In addition, among the service factors, the coding performance factor, network performance factor, and terminal performance factor are viewed through the application quality (quality in application usage such as playback stop, playback start wait, bit rate fluctuation, etc.) that the user actually experiences. It is thought to affect time.

Therefore, if the relationship between the viewing time and various factors that influence it (service factors, application quality, etc.) is clarified, the video distribution service provider can set controllable conditions (encoding conditions, etc.) among the service factors. Optimal conditions can be selected after quantitatively evaluating the effect of the change on the viewing time, and distribution design based on the viewing time becomes possible.

As prior art, there is a technique (Non-Patent Document 1-7) for evaluating QoE based on media information (bit rate, resolution, frame rate) and buffering information (playback stop count, playback stop time, playback stop interval). ..

In addition, there is a technique (Non-Patent Document 8-11) for clarifying the relationship between viewing time and similar indexes (viewing withdrawal rate) and various factors that influence them. Non-Patent Document 8 describes a playback stop ratio (= playback stop time / session time) (a session is from the start to the end of viewing of a specific content by a specific user) and a playback stop count ratio (= playback stop count /). The relationship between viewing time and application quality indicators such as session time), average bit rate, and frame rate is visualized, and some indicators show the correlation with viewing time.

Similarly, Non-Patent Document 9 also visualizes the relationship between the playback start waiting time and the viewing withdrawal rate (= number of viewing withdrawal sessions / total number of sessions) and the standardized playback stop time (= playback stop time / content length) with the viewing time. , Shows the correlation between the two.

Non-Patent Document 10 describes a model that estimates viewing time based on application quality indicators such as playback stop ratio (= playback stop time / session time), playback stop count ratio (= playback stop count / session time), and average bit rate. I am presenting it.

Non-Patent Document 11 visualizes the relationship between the playback stop ratio (= playback stop time / session time) and the bit rate fluctuation ratio (= bit rate fluctuation cumulative value / viewing time) and the viewing withdrawal rate.

Parametric bitstream-based quality assessment of progressive download and adaptive audiovisual streaming services over reliable transport, Recommendation ITU-T P.1203, Nov. 2016. Parametric bitstream-based quality assessment of progressive download and adaptive audiovisual streaming services over reliable transport --Video quality estimation module, Recommendation ITU-T P.1203.1, Dec. 2016. Parametric bitstream-based quality assessment of progressive download and adaptive audiovisual streaming services over reliable transport --Audio quality estimation module, Recommendation ITU-T P.1203.2, Nov. 2016. Parametric bitstream-based quality assessment of progressive download and adaptive audiovisual streaming services over reliable transport --Quality integration module, Recommendation ITU-T P.1203.3, Dec. 2016. K. Yamagishi and T. Hayashi, "Parametric Quality-Estimation Model for Adaptive-Bitrate-Streaming Services," IEEE Transactions of Multimedia, Vol.19, No.7, pp.1545-1557, Jul. 2017. W. Robitza, M. N. Garcia, and A. Raake, "A Modular HTTP Adaptive Streaming QoE Model --Candidate for ITU-T P.1203 (" P.NATS ")," Proceedings of the 2017 Ninth International Conference on Quality of Multimedia Experience (QoMEX), May 2017. A. Raake, M. N. Garcia, W. Robitza, P. List, S. Goring, B. Feiten, "A Bitstream-based, scalable Video-Quality Model for HTTP Adaptive Streaming: ITU-T P.1203.1," Proceedings of the 2017 Ninth International Conference on Quality of Multimedia Experience (QoMEX), May 2017. F. Dobrian, V. Sekar, A. Awan, I. Stoica, D. Joseph, A. Ganjam, J. Zhan, H. Zhang, "Understanding the impact of video quality on user engagement," in Proc. Of SIGCOMM' 11, 2011. S. S. Krishnan and R. K. Sitaraman, "Video Stream Quality Impacts Viewer Behavior: Inferring Causality Using Quasi-Experimental Designs," Proceedings of the 2012 Internet Measurement Conference, pp.211-224, Nov. 2012. A. Balachandran, V. Sekar, A. Akella, S. Seshan, I. Stoica, and H. Zhang, "Developing a Predictive Model of Quality of Experience for Internet Video," Proceedings of the ACM SIGCOMM 2013 conference, pp.339 -350, Aug. 2013. H. Nam, K. Kim, and H. Schulzrinne, "QoE Matters More Than QoS: Why People Stop Watching Cat Videos," Proceedings of the IEEE INFOCOM 2016 --The 35th Annual IEEE International Conference on Computer Communications, Apr. 2016.

However, Non-Patent Document 1-7 proposes a model for estimating QoE, and does not estimate viewing time.

Non-Patent Documents 8, 9 and 11 describe viewing time and similar indexes (viewing withdrawal rate) and application quality indexes (playback stop ratio, playback stop count ratio, average bit rate, frame rate, playback start waiting time) related to the index. , Standardized playback stop time, bit rate fluctuation rate, etc.), but we have not yet presented a model that estimates the viewing time based on the application quality index.

Further, Non-Patent Document 10 presents a model for estimating viewing time based on an application quality index (playback stop ratio, playback stop count ratio, average bit rate, etc.), but this has two problems. ..

First, the viewing time estimation model presented is based on a decision tree (binary tree) model, and because the depth of the tree is shallow, the viewing time estimates are concentrated on several types of discrete values. It will not be put to practical use.

Secondly, although this is a problem common to Non-Patent Document 8, application quality indicators such as playback stop ratio (= playback stop time / session time) and playback stop count ratio (= playback stop count / session time) and viewing Naturally, the time correlation is high (in the negative direction), and a viewing time estimation model based on these indicators does not make sense. This is because the session time in the denominator of the playback stop ratio (= playback stop time / session time) and the playback stop count ratio (= playback stop count / session time) has a very high (positive) correlation with the viewing time. Since it is an index, even if there is almost no correlation between the playback stop time or the number of playback stops and the viewing time, it is as if the reciprocal of the session time had an effect (in the negative direction) between the playback stop ratio or the playback stop count ratio and the viewing time. This is because it looks as if there is a correlation. Therefore, the viewing time estimation based on these indexes means that the viewing time is estimated based on the reciprocal of the session time, and is not practically used either.

The present invention has been made in view of the above points, and an object of the present invention is to support selection of coding conditions for increasing viewing time in adaptive bit rate video distribution.

Therefore, in order to solve the above problem, the distribution design support method is based on the actual value of the viewing time of each of the first plurality of sessions related to the adaptive bit rate video distribution and the actual value of the application quality index of each session. The difference between the first estimation procedure for estimating the relationship between the viewing time or viewing withdrawal rate and the application quality index by the mathematical model, the plurality of coding conditions for adaptive bit rate video distribution, and the first plurality of sessions. A second estimation procedure for estimating the application quality index of each session and a second estimation for each coding condition based on the actual value of the time-series throughput of each of the second plurality of sessions regardless. For each coding condition described in the procedure, the application quality index estimated in the second estimation procedure under the coding condition is applied to the relationship estimated in the first estimation procedure, and the second plurality of said. The computer performs a third estimation procedure for estimating the average viewing time of the session.

It is possible to support the selection of coding conditions for increasing the viewing time in adaptive bit rate video distribution.

It is a figure which shows the hardware configuration example of the delivery design support apparatus 10 in 1st Embodiment. It is a figure which shows the functional configuration example of the delivery design support apparatus 10 in 1st Embodiment. It is a flowchart for demonstrating an example of the processing procedure executed by distribution design support apparatus 10 in 1st Embodiment. It is a figure which shows the functional configuration example of the delivery design support apparatus 10 in the 2nd Embodiment. It is a flowchart for demonstrating an example of the processing procedure executed by the distribution design support apparatus 10 in the 2nd Embodiment. It is a figure which shows the functional configuration example of the delivery design support apparatus 10 in the 3rd Embodiment. It is a flowchart for demonstrating an example of the processing procedure executed by distribution design support apparatus 10 in 3rd Embodiment. It is a figure which shows the functional configuration example of the delivery design support apparatus 10 in 4th Embodiment. It is a flowchart for demonstrating an example of the processing procedure executed by distribution design support apparatus 10 in 4th Embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a hardware configuration example of the distribution design support device 10 according to the first embodiment. The distribution design support device 10 of FIG. 1 has a drive device 100, an auxiliary storage device 102, a memory device 103, a CPU 104, an interface device 105, and the like, which are connected to each other by a bus B, respectively.

The program that realizes the processing in the distribution design support device 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed in the auxiliary storage device 102 from the recording medium 101 via the drive device 100. However, the program does not necessarily have to be installed from the recording medium 101, and may be downloaded from another computer via the network. The auxiliary storage device 102 stores the installed program and also stores necessary files, data, and the like.

The memory device 103 reads and stores the program from the auxiliary storage device 102 when the program is instructed to start. The CPU 104 executes the function related to the distribution design support device 10 according to the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

FIG. 2 is a diagram showing a functional configuration example of the distribution design support device 10 according to the first embodiment. As shown in FIG. 2, the distribution design support device 10 supports the distribution design related to the adaptive bit rate video distribution service, so that the viewing withdrawal rate function estimation unit 11, the application quality index estimation unit 12, and the coding condition optimization It has a part 13 and the like. Each of these parts is realized by a process of causing the CPU 104 to execute one or more programs installed in the distribution design support device 10.

The viewing withdrawal rate function estimation unit 11 inputs "viewing time of all sessions occurring in the past fixed period" and "application quality index of all the sessions", and determines the relationship between the viewing time and the application quality index and the viewing withdrawal rate. The "viewing withdrawal rate function" shown is output. The viewing withdrawal rate is the probability that viewing will be withdrawn (finished) at the viewing time for a session in which viewing has been continued until the viewing time for any viewing time.

"Viewing time of all sessions that occurred in the past fixed period" means all the sessions that occurred in the past fixed period (arbitrary set period such as 1 hour, 1 day, 1 week, etc.) (specific by a specific user). The viewing time (from the start of viewing the content to the end of viewing). It does not matter whether the users and contents of each session are different.

The "application quality index of all the sessions" is the actual result of the application quality index (playback stop time, bit rate fluctuation, etc.) of each session corresponding to the viewing time of the input "viewing time of all sessions occurring in the past fixed period". The value.

With respect to this input, the viewing withdrawal rate function estimation unit 11 estimates a "viewing withdrawal rate function" which is a function indicating the viewing withdrawal rate with respect to an arbitrary viewing time and application quality index. The "viewing withdrawal rate function" corresponds to the hazard function of the proportional hazard model in survival time analysis, and is the viewing withdrawal rate for the same viewing time, taking into account the influence of application quality indicators such as playback stop time. But it varies depending on the session.

Specifically, the viewing withdrawal rate at the viewing time t of the session i, that is, the hazard function according to the viewing time of the session i (under the premise that the session i continued viewing until the viewing time t, the viewing time t The probability of withdrawal from viewing) h ( _zi , t) is expressed by the following equation (1).

ここで、ｚ_ｉは、セッションｉのアプリケーション品質指標ベクトル、γは、アプリケーション品質指標ベクトルｚ_ｉに対応する係数ベクトル、ｈ_０（ｔ）は、基準品質（アプリケーション品質指標ベクトルがゼロベクトルの品質）の下で視聴する基準セッションの視聴時間ｔにおける視聴離脱率である。すなわち、「視聴離脱率関数」は、セッションごとに異なるアプリケーション品質指標の影響を基準セッションの視聴離脱率に乗ずる形で個々のセッションの視聴離脱率を表現するモデルである。

Here, z _i is the application quality index vector of the session i, γ is the coefficient vector corresponding to the application quality index vector z _i , and h ₀ (t) is the standard quality (the quality of the application quality index vector is zero vector). It is a viewing withdrawal rate at the viewing time t of the reference session to be viewed under. That is, the "viewing withdrawal rate function" is a model that expresses the viewing withdrawal rate of each session by multiplying the viewing withdrawal rate of the reference session by the influence of the application quality index different for each session.

The application quality index vector _zi in the equation (1) is a vector consisting of variables such as a playback start waiting time, a playback stop count, an average playback stop time, an average bit rate, a bit rate fluctuation count, and a bit rate fluctuation cumulative value. Is. These variables may be quantitative or qualitative variables. Moreover, even if the original variable is a quantitative variable, it may be converted into a qualitative variable (category variable or dummy variable) as appropriate. For example, when the average bit rate takes a value in the range of 90 kbps to 3500 kbps, the quantitative variable may be input to the viewing exit rate function estimation unit 11, or, for example, the average bit rate is less than 150 kbps, 150 kbps. Divided into 6 categories of less than 300 kbps, 300 kbps or more and less than 450 kbps, 450 kbps or more and less than 750 kbps, 750 kbps or more and less than 2250 kbps, 2250 kbps or more, and the viewing withdrawal rate as a category variable to which the numerical values 1, 2, 3, 4, 5, and 6 are assigned in order. It may be input to the function estimation unit 11. Alternatively, the category variables of 6 categories may be converted into 5 dummy variables as follows and input to the viewing withdrawal rate function estimation unit 11. That is, when the category with an average bit rate of 2250 kbps or higher is used as a reference, a dummy variable indicating whether or not the category corresponds to a category of 150 kbps or less, a dummy variable indicating whether or not the category corresponds to a category of 150 kbps or more and less than 300 kbps, and a dummy variable indicating whether or not the category corresponds to 300 kbps or more and 450 kbps. 5 of a dummy variable indicating whether or not it corresponds to a category less than, a dummy variable indicating whether or not it corresponds to a category of 450 kbps or more and less than 750 kbps, and a dummy variable indicating whether or not it corresponds to a category of 750 kbps or more and less than 2250 kbps. It may be converted into individual dummy variables.

The coefficient vector γ in the equation (1) is the viewing time ti ( _i = 1, 2, ..., N) of N sessions observed in the video viewing experiment of adaptive bit rate video distribution, and each session. Estimated based on the application quality index vector _zi . The estimation is performed by the method of maximizing the partial likelihood expressed by the following equation (2).

ここで、Ｄは、実験で観測されたＮ個のセッションのうち実験終了により視聴終了にならなかったセッション数、ｉ（＝１，２，...，Ｄ）は、実験終了により視聴終了にならなかったセッションのＩＤ、Ｒ（ｔ_ｉ）は、時刻ｔ_ｉにおけるリスク集合、すなわち、Ｎ個のセッションのうち視聴時間がｔ_ｉ以上であったセッションの集合である。

Here, D is the number of sessions that did not end viewing due to the end of the experiment among the N sessions observed in the experiment, and i (= 1, 2, ..., D) ended viewing due to the end of the experiment. The ID and R ( _ti ) of the sessions that did not become are the risk set at the time _ti , that is, the set of the sessions in which the viewing time was _ti or more among the N sessions.

The application quality index estimation unit 12 inputs "coding condition candidate proposal", "time-series throughput of all sessions generated in the past fixed period", and "delivery algorithm", and "in the past fixed period for each coding condition proposal". "Application quality index of all generated sessions" is output.

The "candidate coding condition" is a candidate coding condition that the video distribution service provider considers as a service provision condition for adaptive bit rate video distribution. That is, the "proposal coding condition" is a candidate regarding how many stages of bit rate the video to be placed on the distribution server is encoded and what kind of numerical value is specifically set to the bit rate level of each stage. It's a plan. For example, Candidate 1 has 6 stages of 1 Mbps, 2.5 Mbps, 5 Mbps, 8 Mbps, 16 Mbps and 40 Mbps, and Candidate 2 has 6 stages of 1.5 Mbps, 4 Mbps, 7.5 Mbps, 12 Mbps, 24 Mbps and 60 Mbps. Set.

In addition, "time-series throughput of all sessions that occurred in the past fixed period" was observed in all the multiple sessions that occurred in the past fixed period (arbitrary set period such as 1 hour, 1 day, 1 week). This is the actual value of time series throughput. The past fixed period targeted by the application quality index estimation unit 12 may be the same as or different from the past fixed period targeted by the viewing withdrawal rate function estimation unit 11. That is, each session targeted by the application quality index estimation unit 12 may be the same as or different from each session targeted by the viewing withdrawal rate function estimation unit 11.

In addition, the "delivery algorithm" determines which bit rate level video segment (a video file is divided into small units, also called a chunk) to be received by the terminal for a throughput that fluctuates from moment to moment. It is an algorithm to determine. This algorithm is given in the form of a rule, and is conditional depending on, for example, whether a segment is being downloaded, whether the buffer is abundant or exhausted, or not. For example, if the buffer is neither abundant nor depleted, there is a rule to select the highest bit rate below that value (for example, 75% or less) based on the measured throughput. Another rule is to select the highest bit rate when the buffer is abundant (for example, 20 seconds or more), and select the lowest bit rate when the buffer is exhausted. In addition, if you are downloading a segment and the download time is significantly longer than the playback time of the segment, we will consider that the throughput has dropped significantly and stop the download, and the segment with the same content will be the highest below the most recently measured throughput. There is also a rule such as re-downloading at the bit rate of.

Based on the above input, the application quality index estimation unit 12 receives and plays back in each session in the past fixed period when the candidate coding conditions and the time-series throughput of all sessions are applied to the adaptive bit rate video distribution. The bit rate level at each time of reception or playback of the video segment to be performed and the application quality index (playback stop time, bit rate fluctuation, etc.) observed as a result are estimated by simulation based on the distribution algorithm.

An existing tool such as Savere (https://github.com/UMass-LIDS/sabre/) is used as a simulation tool that outputs the application quality index for the input of coding conditions, distribution algorithm, and time series throughput. can do.

The coding condition optimization unit 13 has generated the "viewing withdrawal rate function" which is the output of the viewing withdrawal rate function estimation unit 11 and the "viewing withdrawal rate function" which is the output of the application quality index estimation unit 12 in the past fixed period for each proposed coding condition. "Application quality index for all sessions" is input, and "coding conditions that maximize the average viewing time" are output.

In response to the above input, the coding condition optimization unit 13 responds to each candidate coding condition as shown in the following equation (3), "for each proposal of coding condition, of all sessions that have occurred in the past fixed period. The "viewing withdrawal rate function" is applied to the "application quality index" to estimate the average viewing time E (t) (the average value of the expected viewing time of each session) in all sessions.

ここで、Ｎは、セッション数、ｆ（ｚ_ｉ，ｔ）は、セッションｉの視聴時間が従う確率密度関数（アプリケーション品質指標ベクトルｚ_ｉに依存する）であり、式（１）のハザード関数ｈ（ｚ_ｉ，ｔ）が求まれば確率密度関数ｆ（ｚ_ｉ，ｔ）も求まる。

Here, N is the number of sessions, and f ( _zi , t) is a probability density function (depending on the application quality index vector _zi ) according to the viewing time of the session i, and the hazard function h in the equation (1). If ( _zi , t) is obtained, the probability density function f ( _zi , t) is also obtained.

The coding condition optimization unit 13 estimates the average viewing time E (t) in all sessions, and then derives the coding condition that maximizes the average viewing time E (t) as the optimum solution.

Hereinafter, the processing procedure executed by the distribution design support device 10 will be described. FIG. 3 is a flowchart for explaining an example of the processing procedure executed by the distribution design support device 10 in the first embodiment.

In step S101, the viewing withdrawal rate function estimation unit 11 determines the viewing time of all the sessions that have occurred in the past fixed period in session units (sessions) and the actual value of the application quality index in session units (sessions). input. The viewing time for each session and the actual value of the application quality index may be stored in, for example, the auxiliary storage device 102 or the like.

Subsequently, the viewing withdrawal rate function estimation unit 11 estimates the viewing withdrawal rate function based on the information input in step S101 (S102).

Subsequently, the application quality index estimation unit 12 sets M coding condition candidate proposals and a distribution algorithm in itself (application quality index estimation unit 12) (S103). The proposed M coding conditions and the distribution algorithm may be stored in the auxiliary storage device 102 in advance, for example.

Subsequently, the application quality index estimation unit 12 inputs the time-series throughput for each session (for each session) for all the plurality of sessions that have occurred in the past fixed period (S104). The past fixed period targeted by the application quality index estimation unit 12 may be the same as or different from the past fixed period targeted by the viewing withdrawal rate function estimation unit 11. That is, each session targeted by the application quality index estimation unit 12 may be the same as or different from each session targeted by the viewing withdrawal rate function estimation unit 11. Further, the time-series throughput for each session may be stored in the auxiliary storage device 102 in advance, for example.

Subsequently, the application quality index estimation unit 12 substitutes 1 for the variable i (S105). The variable i is an ID for identifying the coding condition of each proposed coding condition. It is assumed that the coding condition ID of each proposed coding condition is an integer from 1 to M. Subsequently, steps S107 to S109 are executed for each coding condition (S106).

In step S107, the application quality index estimation unit 12 executes a simulation in which the coding condition (i) and the time-series throughput of the session are applied to the adaptive bit rate video distribution on a session-by-session basis based on the distribution algorithm. Estimate the application quality index for the session.

Subsequently, the coding condition optimization unit 13 estimates the average viewing time in all sessions by applying the estimated application quality index for each session to the equation (3) (S108). Subsequently, the coding condition optimization unit 13 adds 1 to i (S109) and returns to step S106.

When steps S107 to S109 are executed for all (M) coding condition candidate proposals (No in S106), the coding condition optimization unit 13 selects coding condition candidate proposals that maximize the average viewing time. It is specified and the coding condition is output (S110). That is, the proposed coding condition candidate when the average viewing time estimated in step S108 is the maximum is output. However, the coding condition optimization unit 13 may sort the M coding condition candidate proposals in descending order of the average viewing time estimated in step S108, and output the sorting result.

As described above, according to the first embodiment, it is possible to construct a mathematical model that estimates the viewing time based on various factors related to the viewing time, and maximizes the average viewing time based on the model. Delivery design (optimization of coding conditions) can be performed. Specifically, by estimating the average viewing time for each set value of the coding condition, the condition for maximizing the average viewing time is derived from the candidate coding conditions. Therefore, it is possible to support the selection of coding conditions (distribution design) for increasing the viewing time in adaptive bit rate video distribution. As a result, it is possible to support the optimization of the coding conditions.

Next, the second embodiment will be described. The second embodiment will explain the differences from the first embodiment. The points not particularly mentioned in the second embodiment may be the same as those in the first embodiment.

FIG. 4 is a diagram showing a functional configuration example of the distribution design support device 10 according to the second embodiment. In FIG. 4, the same or corresponding parts as those in FIG. 2 are designated by the same reference numerals, and the description thereof will be omitted. As shown in FIG. 4, the function of the distribution design support device 10 is the same as that of the first embodiment, but there are differences in input / output.

In FIG. 4, "user attributes of all sessions that have occurred in the past fixed period" are added to the input of the viewing withdrawal rate function estimation unit 11. This is the actual value of the user attribute regarding the viewer of the video delivered for each session corresponding to the "viewing time of all sessions generated in the past fixed period" which is also the input of the viewing withdrawal rate function estimation unit 11.

With the addition of "user attributes of all sessions that have occurred in the past fixed period" to the input of the viewing withdrawal rate function estimation unit 11, the hazard function represented by the equation (1) in the first embodiment is , In the second embodiment, it is changed to the following equation (4).

ここで、ｘ_ｉ、ｚ_ｉは、セッションｉのユーザ属性ベクトル、アプリケーション品質指標ベクトル、α、γは、それぞれユーザ属性ベクトルｘ_ｉ、アプリケーション品質指標ベクトルｚ_ｉに対応する係数ベクトル、ｈ_０（ｔ）は、基準ユーザ（ユーザ属性ベクトルがゼロベクトルのユーザ）が基準品質（アプリケーション品質指標ベクトルがゼロベクトルの品質）の下で視聴する基準セッションの視聴時間ｔにおける視聴離脱率である。すなわち、第２の実施の形態において、「視聴離脱率関数」は、セッションごとに異なるユーザ属性、アプリケーション品質指標の影響を基準セッションの視聴離脱率に乗ずる形で個々のセッションの視聴離脱率を表現するモデルである。

Here, x _i and z _i are the user attribute vector of the session i and the application quality index vector, and α and γ are the user attribute vector x _i and the coefficient vector corresponding to the application quality index vector z _i , h ₀ (t). ) Is the viewing withdrawal rate at the viewing time t of the reference session that the reference user (user whose user attribute vector is a zero vector) views under the reference quality (quality of the application quality index vector is a zero vector). That is, in the second embodiment, the "viewing withdrawal rate function" expresses the viewing withdrawal rate of each session by multiplying the viewing withdrawal rate of the reference session by the influence of the user attribute and the application quality index different for each session. It is a model to do.

The user attribute vector x _i in the equation (4) is, for example, a vector composed of variables such as gender, age, and user ID. These variables may be quantitative or qualitative variables. Moreover, even if the original variable is a quantitative variable, it may be converted into a qualitative variable (category variable or dummy variable) as appropriate.

The coefficient vectors α and γ in the equation (4) are the viewing time ti ( _i = 1, 2, ..., N) of N sessions observed in the video viewing experiment of the adaptive bit rate video distribution, and each session. Estimated based on the user attribute vector x _i and the application quality index vector z _i . The estimation is performed by the method of maximizing the partial likelihood represented by the following equation (5).

ここで、Ｄは、実験で観測されたＮ個のセッションのうち実験終了により打ち切られなかったセッション数、ｉ（＝１，２，...，Ｄ）は、実験終了により打ち切られなかったセッションのＩＤ、Ｒ（ｔ_ｉ）は、時刻ｔ_ｉにおけるリスク集合、すなわちＮ個のセッションのうち視聴時間がｔ_ｉ以上であったセッションの集合である。

Here, D is the number of sessions not terminated by the end of the experiment among the N sessions observed in the experiment, and i (= 1, 2, ..., D) is the session not terminated by the end of the experiment. ID, R ( _ti ) is a risk set at time _ti , that is, a set of sessions in which the viewing time is _ti or more among N sessions.

Further, as shown in FIG. 4, in the second embodiment, "user attributes corresponding to the application quality index of the left period and all the sessions on the left" are added to the input of the coding condition optimization unit 13. Here, the period on the left is the same period as the past fixed period (arbitrary set period such as 1 hour, 1 day, 1 week, etc.) targeted when the application quality index estimation unit 12 estimates the application quality index. .. In addition, all sessions on the left are all sessions that occurred during the same period.

With the addition of "user attributes corresponding to the application quality index of the left period and all sessions on the left" to the input of the coding condition optimization unit 13, it is expressed by the equation (3) in the first embodiment. The average viewing time E (t) is changed to the following equation (6) in the second embodiment.

ここで、Ｎは、セッション数、ｆ（ｘ_ｉ，ｚ_ｉ，ｔ）は、セッションｉの視聴時間が従う確率密度関数（ユーザ属性ベクトルｘ_ｉおよびアプリケーション品質指標ベクトルｚ_ｉに依存する）であり、式（４）のハザード関数ｈ（ｘ_ｉ，ｚ_ｉ，ｔ）が求まれば確率密度関数ｆ（ｘ_ｉ，ｚ_ｉ，ｔ）も求まる。

Here, N is the number of sessions, and f (x _i , z _i , t) is a probability density function (depending on the user attribute vector x _i and the application quality index vector z _i ) according to the viewing time of the session i. , If the hazard function h (x _i , z _i , t) in Eq. (4) is obtained, the probability density function f (x _i , z _i , t) is also obtained.

FIG. 5 is a flowchart for explaining an example of the processing procedure executed by the distribution design support device 10 in the second embodiment. In FIG. 5, the same steps as those in FIG. 3 are assigned the same step numbers, and the description thereof will be omitted. As shown in FIG. 5, the flow of processing executed by the distribution design support device 10 is basically the same as that of the first embodiment, but there are some differences between step S101 and step S104. Step S101 is replaced by step S101a, and step S104 is replaced by step S104a.

Specifically, "actual value of user attribute for each session that has occurred in the past fixed period" is added to the input of step S101a. This is used as an input when the viewing withdrawal rate function estimation unit 11 estimates the viewing withdrawal rate function in step S102, and is also the input of step S101a, "viewing time in units of sessions generated in the past fixed period". , And the application quality index "must be from the same session for the same period.

Further, "actual value of user attribute for each session generated in the past fixed period" is added to the input of step S104a. This is used as an input when the coding condition optimization unit 13 estimates the average viewing time in all sessions in step S108, and is also the input of step S108, "estimation of application quality index for each session". Must be from the same session for the same period as the "value".

As described above, according to the second embodiment, it is possible to further construct a mathematical model for estimating the viewing time in consideration of the user attributes of each session, and the average viewing time is maximized based on the model. It is possible to perform distribution design (optimization of coding conditions) so as to be.

Next, a third embodiment will be described. The third embodiment will explain the differences from the first embodiment. The points not particularly mentioned in the third embodiment may be the same as those in the first embodiment.

FIG. 6 is a diagram showing a functional configuration example of the distribution design support device 10 according to the third embodiment. In FIG. 6, the same or corresponding parts as those in FIG. 2 are designated by the same reference numerals, and the description thereof will be omitted. As shown in FIG. 6, the function of the distribution design support device 10 is the same as that of the first embodiment, but there are differences in input / output.

In FIG. 6, "content attributes of all sessions that have occurred in the past fixed period" are added to the input of the viewing withdrawal rate function estimation unit 11. This is the actual value of the content attribute related to the content of the video delivered in each session corresponding to the "viewing time of all sessions generated in the past fixed period" which is also input by the viewing withdrawal rate function estimation unit 11.

With the addition of "content attributes of all sessions that occurred in the past fixed period" to the input of the viewing withdrawal rate function estimation unit 11, the hazard function represented by the equation (1) in the first embodiment is , In the third embodiment, it is changed to the following equation (7).

ここで、ｙ_ｉ、ｚ_ｉは、セッションｉのコンテンツ属性ベクトル、アプリケーション品質指標ベクトル、β、γは、それぞれコンテンツ属性ベクトルｙ_ｉ、アプリケーション品質指標ベクトルｚ_ｉに対応する係数ベクトル、ｈ_０（ｔ）は、基準コンテンツ（コンテンツ属性ベクトルがゼロベクトルのコンテンツ）を基準品質（アプリケーション品質指標ベクトルがゼロベクトルの品質）の下で視聴する基準セッションの視聴時間ｔにおける視聴離脱率である。すなわち、第３の実施の形態において、「視聴離脱率関数」は、セッションごとに異なるコンテンツ属性、アプリケーション品質指標の影響を基準セッションの視聴離脱率に乗ずる形で個々のセッションの視聴離脱率を表現するモデルである。

Here, y _i and z _i are the content attribute vector and the application quality index vector of the session i, and β and γ are the content attribute vector y _i and the coefficient vector corresponding to the application quality index vector z _i , h ₀ (t). ) Is the viewing withdrawal rate at the viewing time t of the reference session for viewing the reference content (content whose content attribute vector is a zero vector) under the reference quality (quality whose application quality index vector is a zero vector). That is, in the third embodiment, the "viewing withdrawal rate function" expresses the viewing withdrawal rate of each session by multiplying the viewing withdrawal rate of the reference session by multiplying the influence of the content attribute and the application quality index different for each session. It is a model to do.

The content attribute vector y _i in the equation (7) is, for example, a vector composed of variables such as a content length, a genre, a ranking, and a content ID. These variables may be quantitative or qualitative variables. Moreover, even if the original variable is a quantitative variable, it may be converted into a qualitative variable (category variable or dummy variable) as appropriate.

The coefficient vectors β and γ in the equation (7) are the viewing time ti ( _i = 1, 2, ..., N) of N sessions observed in the video viewing experiment of the adaptive bit rate video distribution, and each session. It is estimated based on the content attribute vector y _i and the application quality index vector z _i . The estimation is performed by the method of maximizing the partial likelihood expressed by the following equation (8).

ここで、Ｄは、実験で観測されたＮ個のセッションのうち実験終了により打ち切られなかったセッション数、ｉ（＝１，２，...，Ｄ）は、実験終了により打ち切られなかったセッションのＩＤ、Ｒ（ｔ_ｉ）は、時刻ｔ_ｉにおけるリスク集合、すなわちＮ個のセッションのうち視聴時間がｔ_ｉ以上であったセッションの集合である。

Here, D is the number of sessions not terminated by the end of the experiment among the N sessions observed in the experiment, and i (= 1, 2, ..., D) is the session not terminated by the end of the experiment. ID, R ( _ti ) is a risk set at time _ti , that is, a set of sessions in which the viewing time is _ti or more among N sessions.

Further, as shown in FIG. 6, in the third embodiment, "content attributes corresponding to the application quality index of the left period and all the sessions on the left" are added to the input of the coding condition optimization unit 13. Here, the period on the left is the same period as the past fixed period (arbitrary set period such as 1 hour, 1 day, 1 week, etc.) targeted when the application quality index estimation unit 12 estimates the application quality index. .. In addition, all sessions on the left are all sessions that occurred during the same period.

With the addition of "content attributes corresponding to the application quality index of the left period and all sessions on the left" to the input of the coding condition optimization unit 13, it is expressed by the equation (3) in the first embodiment. The average viewing time E (t) is changed to the following equation (9) in the third embodiment.

ここで、Ｎは、セッション数、ｆ（ｙ_ｉ，ｚ_ｉ，ｔ）は、セッションｉの視聴時間が従う確率密度関数（コンテンツ属性ベクトルｙ_ｉおよびアプリケーション品質指標ベクトルｚ_ｉに依存する）であり、式（７）のハザード関数ｈ（ｙ_ｉ，ｚ_ｉ，ｔ）が求まれば確率密度関数ｆ（ｙ_ｉ，ｚ_ｉ，ｔ）も求まる。

Here, N is the number of sessions, and f (y _i , z _i , t) is a probability density function (depending on the content attribute vector y _i and the application quality index vector z _i ) according to the viewing time of the session i. , If the hazard function h (y _i , z _i , t) of the equation (7) is obtained, the probability density function f (y _i , z _i , t) is also obtained.

FIG. 7 is a flowchart for explaining an example of the processing procedure executed by the distribution design support device 10 in the third embodiment. In FIG. 7, the same steps as those in FIG. 3 are assigned the same step numbers, and the description thereof will be omitted. As shown in FIG. 7, the flow of processing executed by the distribution design support device 10 is basically the same as that of the first embodiment, but there are some differences between step S101 and step S104. Step S101 is replaced by step S101b, and step S104 is replaced by step S104b.

Specifically, "actual value of content attribute for each session generated in the past fixed period" is added to the input of step S101b. This is used as an input when the viewing withdrawal rate function estimation unit 11 estimates the viewing withdrawal rate function in step S102, and is also the input of step S101b, "viewing time in units of sessions generated in the past fixed period". , And the application quality index "must be from the same session for the same period.

Further, the "actual value of the content attribute for each session that has occurred in the past fixed period" is also added to the input of step S104b. This is used as an input when the coding condition optimization unit 13 estimates the average viewing time in all sessions in step S108, and is also the input of step S108, "estimation of application quality index for each session". Must be from the same session for the same period as the "value".

As described above, according to the third embodiment, it is possible to further construct a mathematical model for estimating the viewing time in consideration of the content attributes of each session, and the average viewing time is maximized based on the model. It is possible to perform a distribution design (optimization of coding conditions) that makes it possible.

In the above, three embodiments based on the proportional hazard model (when considering only the application quality index as a factor affecting the viewing withdrawal rate, when considering the application quality index and the user attribute, the application quality index and the content attribute are used. (When considered) has been described, but as another embodiment, a case where all the factors of the application quality index, the user attribute, and the content attribute are considered can be mentioned.

Next, a fourth embodiment will be described. The fourth embodiment will explain the differences from the first embodiment. The points not particularly mentioned in the fourth embodiment may be the same as those in the first embodiment.

FIG. 8 is a diagram showing a functional configuration example of the distribution design support device 10 according to the fourth embodiment. In FIG. 8, the same or corresponding parts as those in FIG. 2 are designated by the same reference numerals, and the description thereof will be omitted. As shown in FIG. 8, in the fourth embodiment, the distribution design support device 10 has a viewing time function estimation unit 14 instead of the viewing withdrawal rate function estimation unit 11. The viewing time function estimation unit 14 is realized by a process of causing the CPU 104 to execute one or more programs installed in the distribution design support device 10.

The viewing time function estimation unit 14 inputs "viewing time of all sessions that have occurred in the past fixed period" and "application quality index of all the sessions", and "viewing time function" showing the relationship between the application quality index and viewing time. Is output.

The "viewing time function" follows the acceleration model in the survival time analysis, and the random variable t representing the viewing time of the session i is expressed by the following equation (10).

ここで、ｚ_ｉは、セッションｉのアプリケーション品質指標ベクトル、γは、アプリケーション品質指標ベクトルｚ_ｉに対応する係数ベクトル、ｔ_０は、基準品質（アプリケーション品質指標ベクトルがゼロベクトルの品質）の下で視聴する基準セッションの視聴時間を表す確率変数である。すなわち、「視聴時間関数」は、セッションごとに異なるアプリケーション品質指標の影響を基準セッションの視聴時間に乗ずる形で個々のセッションの視聴時間を表現するモデルである。

Here, z _i is the application quality index vector of the session i, γ is the coefficient vector corresponding to the application quality index vector z _i , and t ₀ is the quality under the reference quality (the quality of the application quality index vector is zero vector). It is a probability variable that represents the viewing time of the reference session to be viewed. That is, the "viewing time function" is a model that expresses the viewing time of each session by multiplying the viewing time of the reference session by the influence of the application quality index that differs for each session.

In the case of the accelerated model, unlike the proportional hazard model, it is necessary to parametrically assume the probability density function f ( _zi , t) that follows the viewing time of session i, and it is an exponential distribution, a Weibull distribution, a lognormal distribution, and a loglogistic distribution. It is possible to assume such as as a distribution. For example, the probability density function f ( _zi , t) assuming the Weibull distribution is expressed by the following equation (11).

式（１０）および式（１１）における係数ベクトルγは、アダプティブビットレート映像配信の映像視聴実験で観測されたＮ個のセッションの視聴時間ｔ_ｉ（ｉ＝１，２，...，Ｎ）、及び各セッションのアプリケーション品質指標ベクトルｚ_ｉに基づき推定される。推定は、以下の式（１２）により表される尤度を最大化する方法により行われる。

The coefficient vector γ in the equations (10) and (11) is the viewing time ti ( _i = 1, 2, ..., N) of N sessions observed in the video viewing experiment of adaptive bit rate video distribution. , And the application quality index vector _zi for each session. The estimation is performed by the method of maximizing the likelihood represented by the following equation (12).

ここで、ｆ（ｚ_ｉ，ｔ_ｉ）は、式（１１）により表される確率密度関数にセッションｉの視聴時間ｔ_ｉとアプリケーション品質指標ベクトルｚ_ｉを代入したもの、Ｓ（ｚ_ｉ，ｔ_ｉ）は、確率密度関数ｆ（ｚ_ｉ，ｔ_ｉ）に対応する生存関数（当該視聴時間まで視聴を継続している確率を表す関数）にセッションｉの視聴時間ｔ_ｉとアプリケーション品質指標ベクトルｚ_ｉを代入したものである。また、δ_ｉは、実験終了により視聴を打ち切られた場合に０、打ち切られていない場合に１をとる０－１変数である。すなわち、視聴時間ｔ_ｉで視聴を打ち切られたセッションについては視聴時間がｔ_ｉ以上である確率を、視聴時間ｔ_ｉでユーザの自由意思により視聴を終了したセッションについては視聴時間がｔ_ｉである確率をすべてのセッションについて掛け合わせた尤度を最大化する方法により係数ベクトルγは推定される。

Here, f ( _zi , ti) is obtained by substituting the viewing time ti of the session _i and the application quality index vector _zi into the probability density function expressed by the equation (11), S ( _zi , t ₎ . _i ) is a survival function (a function representing the probability that viewing is continued until the viewing time) corresponding to the probability density function f ( _zi , ti ₎ , the viewing time ti of session _i , and the application quality index vector z. It is the one in which _i is substituted. Further, δ _i is a 0-1 variable that takes 0 when the viewing is stopped due to the end of the experiment and 1 when the viewing is not stopped. That is, the probability that the viewing time is _ti or more for the session in which the viewing is terminated at the viewing time _ti , and the viewing time is _ti for the session in which the viewing is terminated at the user's free will at the viewing time _ti . The coefficient vector γ is estimated by a method that maximizes the likelihood of multiplying the probabilities for all sessions.

FIG. 9 is a flowchart for explaining an example of the processing procedure executed by the distribution design support device 10 in the fourth embodiment. In FIG. 9, the same steps as those in FIG. 3 are assigned the same step numbers, and the description thereof will be omitted. As shown in FIG. 9, the flow of processing executed by the distribution design support device 10 is basically the same as that of the first embodiment, but there are some differences between step S102 and step S108. Step S102 is replaced by step S102a, and step S108 is replaced by step S108a.

Specifically, in step S102a, the viewing time function estimation unit 14 estimates the viewing time function. Further, in step S108a, the coding condition optimization unit 13 estimates the average viewing time based on the equation (3), and the probability density function f ( _zi , t) in the equation (3) at this time is the equation (11). ).

In the above, one embodiment based on the acceleration model (when only the application quality index is considered as a factor affecting the viewing withdrawal rate) has been described, but as another embodiment, the same as in the case of the proportional hazard model. When considering the application quality index and the user attribute, when considering the application quality index and the content attribute, and when considering all the factors of the application quality index, the user attribute, and the content attribute can be mentioned. That is, at least one of the second and third embodiments may be combined with respect to the fourth embodiment.

In the above, examples of proportional hazard model and acceleration model are given, but other models for estimating viewing time include linear regression, nonlinear regression, regression tree, random forest, support vector regression, neural network, and Bayesian network. Various regression system models may be used.

In the present embodiment, the viewing withdrawal rate function estimation unit 11 or the viewing time function estimation unit 14 is an example of the first estimation unit. The application quality index estimation unit 12 is an example of the second estimation unit. The coding condition optimization unit 13 is an example of a third estimation unit.

Although the embodiments of the present invention have been described in detail above, the present invention is not limited to such specific embodiments, and various aspects are within the scope of the gist of the present invention described in the claims. It can be transformed and changed.

10 Distribution design support device 11 Viewing exit rate function estimation unit 12 Application quality index estimation unit 13 Coding condition optimization unit 14 Viewing time function estimation unit 100 Drive device 101 Recording medium 102 Auxiliary storage device 103 Memory device 104 CPU
105 Interface device B Bus

Claims (7)

Based on the actual value of the viewing time of each of the first plurality of sessions related to adaptive bit rate video distribution and the actual value of the application quality index of each session, the relationship between the viewing time or viewing withdrawal rate and the application quality index is determined. The first estimation procedure to estimate by a mathematical model and
The coding condition is based on the plurality of coding conditions relating to the adaptive bit rate video distribution and the actual value of the time series throughput of each of the second plurality of sessions regardless of the difference from the first plurality of sessions. For each, a second estimation procedure for estimating the application quality index for each session,
For each coding condition described in the second estimation procedure, the application quality index estimated in the second estimation procedure under the coding condition is applied to the relationship estimated in the first estimation procedure. A third estimation procedure for estimating the average viewing time of two multiple sessions, and
A delivery design support method characterized by a computer running. The first estimation procedure further estimates the relationship between the viewing time or the viewing withdrawal rate and the user attribute based on the actual value of the user attribute.
The delivery design support method according to claim 1, wherein the delivery design is supported. The first estimation procedure further estimates the relationship between the viewing time or the viewing withdrawal rate and the content attribute based on the actual value of the content attribute.
The delivery design support method according to claim 1 or 2, wherein the delivery design is supported. The third estimation procedure identifies the coding condition that maximizes the average viewing time.
The delivery design support method according to any one of claims 1 to 3, wherein the delivery design is supported. Based on the actual value of the viewing time of each of the first plurality of sessions related to adaptive bit rate video distribution and the actual value of the application quality index of each session, the relationship between the viewing time or viewing withdrawal rate and the application quality index is determined. The first estimation part estimated by the mathematical model and
The coding condition is based on the plurality of coding conditions relating to the adaptive bit rate video distribution and the actual value of the time series throughput of each of the second plurality of sessions regardless of the difference from the first plurality of sessions. For each, a second estimator that estimates the application quality index for each session,
For each coding condition described in the second estimation unit, the application quality index estimated by the second estimation unit under the coding condition is applied to the relationship estimated by the first estimation unit. A third estimation unit that estimates the average viewing time of two multiple sessions,
A delivery design support device characterized by having. The third estimation unit identifies the coding condition that maximizes the average viewing time.
The delivery design support device according to claim 5. A program characterized by causing a computer to execute the distribution design support method according to any one of claims 1 to 4.

Images