JP2011176402A - Content prior distribution method, system and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To increase the ratio of actually viewing contents distributed in advance while reducing the required storage amount of STB. <P>SOLUTION: In the content prior distribution method, a user cluster design part 103 and a distribution schedule design part 104 are provided. The user cluster design part 103 classifies users into a plurality of clusters of close content viewing tendency and sends the classified clusters to the distribution schedule design part 104 for each cycle (one day) of prior distribution. When preparing the BC prior distribution schedule of the respective clusters, the distribution schedule design part 104 suppresses BC prior distribution contents in the respective clusters to only highly popular contents for a smaller number other than the maximum number of contents that can be distributed in advance during one cycle of the BC prior distribution of the respective clusters. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a content pre-distribution method, system, and apparatus for pre-distributing popular content by broadcast from a distribution server to all users periodically via a network.

  Many ISPs (Internet Service Providers) position video-on-demand services (VoD services) as a killer service that differentiates their network services, and high quality produced by content production companies such as movies and TV dramas. Is focusing on popularizing VoD services that distribute video content (rich content).

  In general, viewing of moving image content tends to concentrate at night, and fluctuates greatly with the cycle of a day. In order for ISPs to improve profits and provide VoD services to users at low cost, reducing server load during peak hours is a major issue. In addition, since rich content has a large capacity, it is also important to suppress the influence of traffic generated by distribution on the network (NW). Various methods for reducing server load during peak hours have been studied and put into practical use, but can be classified mainly into MC distribution (multicast distribution) and P2P (Peer to Peer) type distribution.

  MC distribution is to distribute to a plurality of users who requested the same content in a single distribution session using the MC distribution tree, and a reduction in server load can be expected in proportion to the improvement in user aggregation. . However, in order to increase the degree of user aggregation, it is necessary to wait for distribution without immediately starting distribution in response to a distribution request. For this reason, the immediacy of distribution is impaired, the waiting time from when the user requests distribution until the distribution is actually started increases, and the service quality deteriorates.

  On the other hand, a P2P type distribution system such as BitTorrent (registered trademark) that realizes content distribution by exchanging content held between users is widely spread. Conventionally, it has been used for the purpose of exchanging UGC (User Generated Contents), but in recent years, P2P distribution technology is also used as a means for distributing rich contents. However, there is a possibility that the user may leave the system in the middle of distribution, and it is necessary to use a measure for continuing distribution such as switching the connection destination to another peer. Moreover, as can be said for MC distribution, in order to support VCR operations such as playback and stop, complicated processing for dynamically switching distribution sessions is required.

  By the way, the increase in the downlink capacity of the access line is remarkable, and an environment in which a sufficiently large band can be used compared with the reproduction rate of rich content is becoming common. For this reason, in addition to the normal delivery that is delivered from the delivery server on demand to the user's delivery request, the unused transmission capacity is utilized, and the content is constantly delivered regardless of the user's delivery request, It is effective to store in an installed STB (set top box). Content distribution performed regardless of the user's distribution request is called pre-distribution.

  That is, when the requested content is already stored in the STB by the prior distribution when the user requests distribution, distribution from the distribution server can be avoided, and the load at the peak of the distribution server and NW can be reduced. Furthermore, when delivering to an STB regardless of a user's delivery request, the ISP can freely set delivery content and delivery timing, so the same content is delivered to all users via BC (broadcast delivery). Thus, it is possible to suppress the load that pre-distribution gives to the distribution server and NW.

  Therefore, the inventors have proposed a BC pre-distribution method as shown in Non-Patent Document 1 as a distribution method in the VoD service. In this BC pre-distribution method, unlike the P2P type distribution, since the content cached in the STB is viewed only by the user who owns each STB, there is no need to give the user an incentive to upload the content to other users. When viewing the pre-distributed content, no traffic is generated in the NW, so that it is possible to expect a reduction in the NW load in addition to the peak server load. Furthermore, it can easily cope with VCR operations.

  In this BC pre-distribution method, the distribution system is composed of a distribution server and distribution NW prepared by an ISP, and a plurality of user-side apparatuses having moving image viewing apparatuses such as STB and TV. The distribution server stores all M contents to be provided. The distribution server immediately unicasts the requested content in response to a distribution request from the user. The user side device reproduces the distributed content while accumulating it in the STB. Further, in addition to on-demand unicast distribution, the distribution server always distributes content in advance, and the STB always stores received content. Regardless of the user's distribution request, the distribution server schedules the content to be pre-distributed on the nth day at the start time (0:00 am) on the nth day. In advance distribution, the same content is broadcast to all users, thereby reducing the influence on the distribution server and NW.

  When the content that the user wants to view has already been pre-distributed and exists in his / her STB, the user can view without requesting the distribution server. Therefore, no new traffic is generated in the NW, and any VCR operation can be easily handled. Therefore, it is possible to greatly reduce the load on the distribution server and the NW by pre-distributing popular content with BC. In general, from the viewpoint of copyright protection, a reproducible period (L day) is set for the content stored in the STB. Therefore, after the content is distributed to the STB, the midnight including the distribution date has passed. Sometimes erase.

  Assume that the downlink capacity of the access line reserved for the VoD service is Ad (bps) and the content playback rate is R (bps). When delivering by unicast from the delivery server, delivery is performed with R, and the delivery rate Rb (bps) of BC delivery is set to Rb = γR with respect to the positive real parameter γ. In order to secure a transmission band for unicast distribution for each user, it is necessary to satisfy γR ≦ (Ad−R). Also, BC pre-delivery is performed with one day as one cycle, and the average value X of the number of contents that can be pre-delivered during one cycle is X = 24 × 3600γ / S (S is content) X is maximized by setting γ = (Ad−R) / R.

The total download time for the content m on the nth day _{, where} Q _{m, n} is the set of distribution requests for the content m that occurred on the 24th hour on the nth day and tq (seconds) is the viewing time of the distribution request q. (TDLT: totaldownloadtime) D _{m, n} (seconds) is

However, the distribution server load is reduced on the nth day as the pre-delivery of the content having a large D _{m, n} preferentially on the nth day. Content to be pre-delivered in the 24th hour of the nth day is determined by the delivery scheduling set at the start time of the nth day. These contents are determined from the nth day to the (n + L-1) th day. Present in the STB for L days. Therefore, it is desirable to preferentially select content having a large sum of D _{m, n} over the L days in the distribution scheduling on the nth day.

However, since D _{m, i} with i ≧ n cannot be obtained at the start time of the nth day, D _{m, n} needs to be estimated. Here, the BC pre-distribution is performed on the n-th day for the content having a large TDLT (D _{m, n-1} ) on the previous day. That is, D _{m, n−1} is used as the estimated value D ′ _{m, n} on the nth day. However, since the BC pre-distributed content from the n-L day to the n-1 day is cached in the STB on the n-th day, the BC advance is made on the (n-L) to (n-1) days. The BC pre-distributed content is selected in descending order of D ′ _{m, n} for the content excluding the distributed content as long as there is free time. Further, the distribution order of the contents is set in descending order of D ′ _{m, n} / Sm with respect to the length Sm (seconds) of the contents _m .

“Broadcast pre-distribution in VoD service”, IEICE Technical Report NS2009-1563.

  However, BC pre-distribution requires a large amount of storage in the STB, and there is a problem that the proportion of pre-distributed content is actually viewed is low. In the evaluation using China Telecom's “PowerInfo VoD system” access log data, the percentage of users who actually watched the BC pre-distributed content is 0.39 when L = 3. When the L = 7, the STB is as low as about 0.51%, and each STB requires a large capacity storage of 2.1 Tbytes when the L = 3 and 4.8 Tbytes when the L = 7. Become.

  The present invention has been made in order to solve such problems. The object of the present invention is to reduce the amount of storage required for the STB and increase the proportion of pre-distributed content that is actually viewed. It is an object of the present invention to provide a content pre-delivery method, system and apparatus capable of performing the above.

  In order to achieve such an object, the present invention periodically provides a content pre-distribution method (system, apparatus) for pre-distributing popular content by broadcast from a distribution server to all users via a network. , Only a few highly popular contents are pre-delivered than the maximum number of contents that can be pre-delivered during one pre-delivery period.

  According to the present invention, when the maximum number that can be pre-distributed during one period of pre-distribution is X, the content of the maximum number X is not pre-distributed but is smaller than the maximum number X. Only highly popular content will be pre-delivered. Thereby, the number of contents delivered in advance is suppressed, and the required storage amount of STB is reduced. In addition, since the pre-distributed content is a small number and highly popular content, the proportion of the content that is actually viewed increases.

  The present invention also provides a period of pre-distribution in a content pre-distribution method (system, apparatus) that pre-distributes popular content by broadcast from a distribution server to all users periodically via a network. Each time the user is classified into a plurality of clusters having a similar viewing tendency of the content, and less popular content than the maximum number of content that can be pre-distributed during one pre-delivery period for each classified cluster Only pre-delivered.

  According to this invention, a user is clustered for every period of prior delivery, and it classify | categorizes into the some cluster with the viewing-and-listening tendency of content close | similar. Here, for example, when the maximum number of the classified clusters that can be pre-distributed in one cycle of the pre-distribution of clusters 1 and 2 and clusters 1 and 2 is X1 and X2, the maximum number X1 and X2 Content is not pre-distributed to the users in clusters 1 and 2, but only highly popular content smaller than the maximum number X1, X2 is pre-distributed to the users in clusters 1 and 2. It will be a thing. Thereby, the number of contents delivered in advance is suppressed, and the required storage amount of STB is reduced. In addition, since the pre-delivered content is a small number and highly popular content for each user group of each cluster, the rate of actual viewing is also increased.

  According to the present invention, since only a small number of highly popular contents are pre-delivered than the maximum number of contents that can be pre-delivered during one period of pre-delivery, each period of pre-delivery The users are classified into a plurality of clusters that are close to the viewing tendency of the content, and only a few highly popular contents than the maximum number of contents that can be pre-distributed during one pre-delivery period for each classified cluster. Is pre-distributed, it is possible to reduce the amount of storage required for the STB and to increase the proportion of the pre-distributed content that is actually viewed.

It is a figure which shows the principal part of one Embodiment of the system to which the content pre-delivery method concerning this invention is applied. It is a flowchart which shows the process in the delivery server in this content pre-delivery system. It is a flowchart which shows the process in the user side apparatus in this content pre-delivery system. It is a flowchart which shows the process in the user cluster design part in this content pre-delivery system. It is a flowchart which shows the process in the delivery schedule design part in this content pre-delivery system. It is a sequence diagram which shows a mode that the immediate unicast delivery and BC predelivery in this content predelivery system are performed. It is a figure where it uses for description for the average TDLT of top x content, and the reduction effect of K and V. It is the figure which collected K and V at the time of using MC delivery and P2P delivery. It is the figure which put together the value of Z, K, V in BC pre-delivery at the time of delivery control to the value of L and r. It is a figure where it uses for description about the characteristic regarding a cache hit rate and a user content viewing tendency. It is the figure which put together K, V, ξ, and η when clustering is performed for each cluster value. It is a figure which shows the time-sequential characteristic in C = 2. It is the figure which plotted the gravity center of each cluster of the 100th day, and the audience rating of each content. It is the figure which plotted the number of users of each cluster of the 100th day. It is the figure which plotted the cache hit rate with respect to the number of clusters. FIG. 10 is a diagram for explaining the effect when the top x contents of the previous day are selected.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a diagram showing a main part of an embodiment of a system to which a content pre-distribution method according to the present invention is applied.

  In FIG. 1, 101 is a distribution server, 102 is a user side device, 103 is a user cluster design unit, 104 is a distribution schedule design unit, 105 is an NW (network), 106 is a distribution server 101, user cluster design unit 103, and This is a content pre-distribution apparatus having the distribution schedule design unit 104 as a constituent element.

  The user side device 102 is a video viewing device provided in each user's home, and has an STB as a content receiving device. The content pre-distribution device 106 is realized by hardware including a processor and a storage device, and a program that realizes various functions as the content pre-distribution device in cooperation with these hardware.

  In this system, when content to be viewed is distributed in advance in the STB of the user-side device 102, the user views the content by playing the locally cached content. When the content to be viewed does not exist in the STB of the user side device 102, the distribution server 101 requests distribution to the distribution server 101, and the distribution server 101 that has received the distribution request immediately unicasts the user side device of the requesting user. The requested content is distributed to 102.

  Further, the distribution server 101 measures the total download time (TDLT) for each content. The user cluster design unit 103 classifies the users into C clusters based on content viewed by each user in the past Tc days at the start of the day (midnight), and the generated cluster is a distribution schedule design unit 104. Notify

  The distribution schedule design unit 104 estimates the TDLT of each content for the next day for each cluster, and designs a BC pre-distribution schedule for each cluster based on the estimation result. The distribution server 101 broadcasts content to the user side devices 102 of the user group of each cluster from midnight to 24r hours (r will be described later) based on the designed BC pre-distribution schedule of each cluster.

  FIG. 2 shows a processing flow of the distribution server 101. The distribution server 101 performs both immediate unicast distribution (FIG. 2A) and BC pre-distribution (FIG. 2B) in response to the distribution request.

  In the immediate unicast distribution, in response to a distribution request from the user (step S101), the requested content is immediately distributed by unicast (step S102). Then, a user access log is recorded (step S103).

  In the BC pre-distribution, according to the distribution schedule of each cluster designed by the distribution schedule design unit 104 (step S201), the contents of each cluster to be pre-delivered BC next are selected (step S202), and the user group of each cluster is selected. BC pre-delivery is performed (step S203). The BC pre-distribution ends when the BC pre-distribution period has elapsed from the start time of each day (YES in step S204).

  FIG. 3 shows a processing flow of the user device 102. When a content viewing request is generated (step S301), it is first checked whether the content has already been BC pre-distributed and exists in the STB (step S302). If it exists in the STB (YES in step S302), the content is reproduced from the STB without receiving a new distribution from the distribution server 101 (step S303). Then, the viewing log is notified to the distribution server 101 (step S304). If it does not exist in the STB (NO in step S302), a distribution request is notified to the distribution server 101 (step S305), and unicast distribution is immediately received from the distribution server 101 (step S306), and playback is performed.

  FIG. 4 shows a processing flow of the user cluster design unit 103. The user cluster design unit 103 collects the history of the content viewed by each user in the past P days from the access log at the start of the day (midnight) (step S401), and uses this collected content viewing history as the viewing history of the collected content. Based on the K-means method (k-means method), the user is clustered into arbitrarily given number C clusters (step S402).

  FIG. 5 shows a processing flow of the distribution schedule design unit 104. For each cluster generated by the user cluster design unit 103, the distribution schedule design unit 104 uses the access log to calculate the total viewing time TDLT for each content of the previous day (step S501), and the TDLT for each content of the next day The estimated value is used (step S502).

  Then, as BC pre-distributed contents in each cluster, contents are allocated in descending order of TDLT value (step S503). At this time, the delivery schedule design unit 104 is more popular than the maximum number of BC pre-delivery contents in each cluster that can be pre-delivered during one cycle (1 day) of BC pre-delivery in each cluster. Limit to content only. Also, content stored in the user-side STB is excluded.

  Then, the delivery schedule design unit 104 sets the delivery order for the BC pre-delivery content of each cluster in descending order of the value obtained by dividing the TDLT estimated value by the content length (corresponding to the estimated value of the number of delivery requests). A BC pre-distribution schedule is set for the cluster (step S504).

  FIG. 6 shows how immediate unicast delivery and BC pre-delivery are performed in this content pre-delivery system. The distribution schedule design unit 104 sets a BC pre-distribution time TB (TB = 24r time) for the distribution server 101 (arrow (1)). When there is a distribution request from the user side device 102 (arrow (2)), the distribution server 101 performs immediate unicast distribution (arrow (3)). The distribution history of each content is notified to the user cluster design unit 103 and the distribution schedule design unit 104 (arrows (4), (5)).

  The user cluster design unit 103 collects the history of content viewed by each user in the past P days at the start of the day (midnight), clusters the users based on the collected content viewing history, The generated cluster is sent to the distribution schedule design unit 104 (arrow (6)).

  The distribution schedule design unit 104 creates a BC pre-distribution schedule for each cluster generated by the user cluster design unit 103 and notifies the distribution server 101 (arrow (7)). The distribution server 101 performs BC pre-distribution of content to the user group of each cluster according to the BC pre-distribution schedule of each cluster from the distribution schedule design unit 104 (arrow (8)).

  Next, details of distribution suppression and user clustering in the BC pre-distribution described above will be described.

[Reduction of distribution (reduction of the number of BC pre-distributed contents)]
In general, user content viewing requests tend to concentrate on a small number of highly popular contents. Therefore, even if BC pre-delivery is always performed on each day and the maximum number of contents are not pre-delivered with BC, only a small number of highly popular contents are pre-delivered with BC. The

  That is, by suppressing the content pre-distributed on each day to 100r% of the maximum value X, the required storage capacity of the STB can be suppressed to 100r%, but the degradation degree of the load reduction effect of the distribution server at this time is 100r % Can be expected to be smaller. Therefore, in FIG. 6, as indicated by the arrow (1), prior to the start of the service, the BC pre-delivery time TB of each day is determined as TB = 24r hours and set in the delivery server 101.

  Next, the effect of intentionally suppressing the content to be pre-delivered to a small number is evaluated. For the evaluation, access log data collected in a commercial VoD service “PowerInfo VoD system” provided by China Telecom is used. This log data includes all the distribution requests of 20,921,657 that occurred during the seven months from June to December 2004. The total number of contents is 6,735 and the total number of users is 37,360. The contents having a length of 45 minutes and 90 minutes each occupy about 40% of the total, and the average content length S is 3,510 seconds. The downlink capacity of the access line is set to Ad = 10 Mbps, and the content reproduction rate is set to R = 2 Mbps. γ = 4, the BC pre-delivery rate is Rb = 8 Mbps, and the average number of BC pre-delivery contents per day is X = 98.46. A load corresponding to Rb / R = 4 unicast distributions is always applied to the distribution server for BC pre-distribution. As a measure for measuring the effectiveness of the content distribution method, the average values K and V of the maximum number of simultaneous sessions Kn of each day n are evaluated for the entire period.

In FIG. 7A, as “Actual” _, the sum of D _{m, n} of the top x pieces of content having a large TDLT (D _{m, n} ) in each day n occupies the sum of D _{m, n} of all contents. The figure which plotted the average value of the ratio over the whole period with respect to x is shown. Further, in FIG. 7 (a), as "Estimate", the sum of TDLT of the n days of top-x content of the n-1_Nichino _{TDLT (D m, n-1} ) (D m, n) is The figure which plotted similarly the average value of the ratio to the sum of _{Dm, n} of all the contents of the nth day is shown.

  In “Actual”, for example, the top 10 content, which is only 0.16% in terms of number, is 18.5% of TDLT, and the top 30 content, which is only 0.48% in terms of number, is TDLT. It can be confirmed that the viewing is concentrated on a small number of contents. Actually, the BC pre-distribution content is selected based on the previous day's TDLT. Therefore, the effect of suppressing BC pre-distribution is equivalent to “Estimate”, and the degree of concentration of the TDLT of the upper content is slightly reduced. The content is 15.6%, the top 30 content is 25.6%, and the majority of the TDLT is also occupied by a small number of content.

When the time TB for performing BC pre-delivery on each day is set to TB = 24 × 3,600 × r (seconds) (the BC pre-delivery time is suppressed to 100 r%), K and V are respectively K (r) and V (R). The relative improvement effects ψ _K (r) and ψ _V (r) of these evaluation measures are respectively expressed as ψ _K (r) = (K (0) −K (r)) / (K (0) −K. (1)), ψ _V (r) = (V (0) −V (r)) / (V (0) −V (1)). However, K (0) and V (0) are K and V when the distribution server immediately performs unicast distribution for all distribution requests without performing BC pre-distribution.

ψ _K (r) and ψ _V (r) are K values when the number of BC pre-distributed contents is suppressed to 100 r% with respect to the reduction amount of K and V when the BC pre-distribution is always performed (r = 1). It corresponds to the ratio of the reduction amount of V. When r <1, how to determine the BC pre-delivery time zone for each day is a problem. From the viewpoint of reducing the peak load of the delivery server, BC pre-delivery may be performed only during off-peak hours. It is valid. However, since the same content is distributed to all users by BC, the number of distribution sessions used for BC pre-distribution is small (Rb / R = 4 in this numerical evaluation example), so the load imposed on the distribution server by BC pre-distribution Is small. Therefore, in this example, BC pre-distribution is performed for a period of TB from 0:00, which is the switching time of the BC pre-distribution schedule.

FIGS. 7B and 7C are diagrams in which ψ _K (r) and ψ _V (r) are plotted against three values of the content storage period L with respect to the pre-delivery time ratio r. ψ _K (r) and ψ _V (r) are curved upwards with respect to r, and exist in the upper left region of the straight line y = x. Therefore, when r is set in the range of 0 <r <1, and the number of BC distribution contents is suppressed, the degree of deterioration of the improvement effect of K and V is smaller than the reduction ratio r of the BC pre-distribution time. Since the necessary storage capacity Z of STB is proportional to r, Z can be reduced while suppressing the performance degradation of the pre-delivery by suppressing the number of BC pre-delivery contents. For example, when r = 0.5 is set when L = 7, a reduction effect on K and V of about 80% can be obtained while Z is halved. Further, when L is determined based on the contract conditions with the content provider, etc., Z can be designed to satisfy desired STB cost requirements by adjusting r.

  Next, in order to confirm the effectiveness of BC pre-distribution in which the number of BC pre-distribution contents is suppressed, performance is compared with the case where MC distribution and P2P distribution are used as typical distribution server load reduction methods in VoD.

  In MC distribution, a waiting period is provided for a user's distribution request, and distribution requests for the same content generated during that period are consolidated into one MC distribution. Here, an upper limit value W is set for the response time of the user, and the content is distributed after W from the time when the request for the content for which there is no other waiting user exists.

  In P2P distribution, the requested content is distributed from other users who are viewing the content to the distribution requesting user. Here, assuming an ideal environment, when each user requests distribution for content m, the product of the number of users receiving distribution of content m and the upload capacity Au of the access line is distributed from other peers. When the product of the number of received users + 1 and the reproduction rate R is larger, the distribution is received from the peer. Otherwise, the distribution is performed from the distribution server. In addition, when this condition is not satisfied when the user finishes the distribution, the distribution server switches the distribution from the distribution server to the peer that has the minimum remaining time until the end of the distribution among the users receiving the distribution from the peer. .

FIG. 8 is a diagram summarizing K and V (10 ⁷ seconds) when the two comparison methods are used. As shown in FIG. 8A, in MC distribution, K and V decrease as the waiting period length W (seconds) increases, but the average response time w (seconds) of the user deteriorates. Further, as shown in FIG. 8B, in P2P distribution, K and V decrease as Au increases, but performance improvement reaches its peak when Au approaches Ad. In this evaluation, performance improvement was not seen when Au was 5 Mbps or more.

FIG. 9 summarizes the values of Z (Gbyte), K, and V (10 ⁷ ) in BC pre-distribution when distribution suppression is performed for three values of L and six values of r. The figure is shown. Even now, it is possible to purchase an HD recorder with a capacity of 1 Tbyte for about 40,000 yen, and the storage cost is expected to continue to be reduced. For example, when L = 7, Z can be suppressed to about 1 Tbyte by setting r = 0.2. From the viewpoint of K and V, MC is distributed with W of about 5 minutes, and Au is about 2 Mbps. It is possible to achieve performance equivalent to that of P2P distribution. On the other hand, all the problems such as coping with VCR operation, quality instability, and coping with user detachment, which are problems in the two comparison methods, can be solved by BC pre-distribution when distribution suppression is performed.

[User Clustering]
Since user preferences vary, it is expected that there are users who prefer to view content of a specific genre that is not popular, rather than popular content that many users view. Therefore, by clustering users who are close to the content viewing tendency and performing BC pre-distribution of content with a large TDLT for each cluster, the proportion of pre-distributed content that is actually viewed (utilization rate of pre-distributed content) η Improvement and further reduction of K and V are expected.

FIG. 10A shows a cumulative distribution of the cache hit rate (CHR) ξ _u of each user u with respect to three values of L. However, ξ _u is defined as the ratio of the contents that can be dealt with by the BC pre-distributed content in the TDLT for all viewing requests of the user u. As this value increases, the effect of reducing K and V increases. The average cache hit rate calculated for all users is 0.281 (L = 1), 0.404 (L = 2), and 0.642 (L = 7). The difference in ξ _u for each user is It can be confirmed that there are many users who are large and small in ξ _u . Since content with a large TDLT targeting all users is pre-delivered to BC, the content viewing tendency of users with a small ξ _u is expected to be significantly different from the viewing tendency of all users.

FIG. 10B shows a diagram in which TDLT (D _u ) over the entire period of each user u is plotted against ξ _u when L = 1. Since D _u is to view large heavy users many contents, high popularity content just not BC prepositioned that are not a result of view many low popular content, xi] _u is relatively of 0.2 to 0.4 being BC prepositioned Concentrate on a small range of values. On the other hand, D _u is the general user of low to medium a result of the tendency of viewing content differs greatly for each user, the user indicating extreme viewing tendency of increased variation in the per-user xi] _u xi] _u is 0 or 1 It can be confirmed that there are many. The relative TDLT (Δ _{u, m, n} ) of each viewing content m on the nth day of the user u, the time when the user u viewed the content m on the nth day is the TDLT on the nth day of the user u Defined by the value divided.

For two users A and B having the same D _u shown in FIG. 10B, Δ _{u, m, n} is the TDLT (D _{m, n} ) for the contents m of all users _n on each day. A diagram plotted in descending order is shown in FIG. User A having a large ξ _u is mainly viewing content pre-delivered to BC because D _{m, n} is large and popular. On the other hand, the user B whose ξ _u is small mainly views niche contents that are not pre-delivered by BC because D _{m, n} is small and low popularity. Thus, there is a case where D _u prone was also view content comparable users differ greatly, such as user B, and xi] _u of different users with multiple users tend to view content, user It may be improved by clustering the set.

Consider that a user set is classified into a plurality of clusters based on the viewing tendency of content during BC pre-distribution scheduling. That is, as shown in FIG. 4, the user is classified into a plurality of clusters from the history information of the viewing content of each user using the user access log. As shown in FIG. 6, the clustering process is performed to generate a cluster to be used for BC pre-delivery on the next day at the end of each day. In this example, a user set is clustered using the k-means method, which is a representative method of division optimization clustering. When there are I clustering objects x _i having J pieces of data x _{i, 1} , x _{i, 2} ,..., X _{i, J} , the k-means method uses the centroid m _c = {m of each cluster c. _{c, 1} , m _{c, 2} ,..., m _{c, J} } and the sum of Euclidean distances between x _i constituting each cluster c,

These I objects are classified into an arbitrarily determined number C clusters so that is minimized. _Xc is a clustering target set classified into cluster c.

The algorithm for the k-means method is shown below.
(1) A clustering target is randomly divided into C clusters to generate an initial cluster.
(2) For each cluster c, centroid _mc

Calculated from
(3) Distance all clustering objects x _i to the center of gravity

Assign to the cluster that minimizes.
(4) When there is a change in the cluster, the cluster is reassigned by repeating (2) and (3).

In this example, the clustering target x _i is each user, and the data set x _{i, j} used for clustering is 1 when the user i views the content j within a certain period Tc, and 0 when the user i does not view it. Consider a binary variable to be classified into arbitrarily defined C clusters. If the number of iterations is regarded as a constant, the calculation amount of the k-means method is O (ICJ). However, if V _i is a set of contents viewed by user i within period T and V ′ _i is its complement,

  In this equation (2),

Is determined for each cluster c regardless of the user x _i, and in each iteration, δ (x _i , m _c ) is calculated with O (| V _i |) for each user x _i and each cluster c. It can be calculated.

  FIG. 10 (d) shows a plot of the cumulative distribution of the number of contents viewed by each user during each day or during the entire period. Although the maximum value of the number of viewed contents for each day is as large as 498, about half of the contents are viewed. The number of viewing contents of most users is 2 or less, and about 90% of the users are 10 or less. Similarly, when Tc is set to the entire period, the maximum value is as large as 5,298, but about half of the users are 141 or less, and about 90% of the users are 301 or less, which is much smaller than the total number of contents. Therefore, the time required for clustering can be significantly reduced by calculating the distance using Equation (2).

Henceforth, in this example, examination is advanced for the case of L = 1. In order to increase the hit rate of BC pre-distributed content and reduce the load on the distribution server, at the start of each day, users who are as close as possible to the viewing content set on that day can be classified into the same cluster. desirable. Therefore, in this example, for the purpose of evaluating the possibility of improving the performance of BC pre-distribution by user clustering, a period Tc for determining x _{i, j} used for clustering at the start time of the nth day is set to the nth day. Cluster user sets as only one day. Then, a value obtained by adding the TDLT (D _{i, j, n} ) for each content j on the nth day of each user i constituting each generated cluster c to each content j.

In the descending order of BC, BC distribution is performed to the users constituting the cluster c. Since the BC pre-distribution content is selected for each cluster, the number of distribution sessions required for BC pre-distribution is C times compared to the case where user clustering is not performed.

  FIG. 11A summarizes K, V, ξ, and η for ideal clustering with respect to seven values (2, 4, 8, 16, 32, 64, and 128) of cluster C. The figure is shown. Here, ξ is an average cache hit rate for all users, and is defined as the ratio of TDLTs for viewing the BC pre-distributed content of all users for all periods. Further, η is the utilization rate of the pre-distributed content as described above.

  In terms of the number of sessions in the unicast distribution flow, Rb / R = 4C sessions are always required for BC pre-distribution, but when calculating K, these 4C sessions are also added. . For comparison, FIG. 11A also shows the results when the same content is pre-delivered to all users without performing clustering (C = 1). However, even in the case of C = 1, BC pre-delivery content was selected in descending order of TDLT of each content of each day.

  As the number of clusters C increases, K and V decrease, ξ and η increase, and all characteristics are improved, and the effect of user clustering in BC pre-distribution can be confirmed. However, the improvement degree of K and V is small compared with the improvement degree of ξ and η. For example, when C = 16, ξ and η increase by about 17% compared to the case of C = 1, but K is only about 2.6% and V is only about 5.4%. This is because ξ and η are greatly increased by user clustering, but because different contents are distributed via BC in each cluster, the distribution session and transfer data amount required for BC pre-distribution increase.

  In particular, since 4C sessions are always required for BC pre-delivery, when C becomes larger than about 16, K worsens with increasing C. On the other hand, V worsens when C becomes larger than about 64. Therefore, when C is increased, it is necessary to reduce the peak load of the distribution server by performing BC pre-distribution while avoiding the peak time. By suppressing the number of BC pre-distributed content, it is possible to effectively reduce K and V with a small amount of STB storage capacity, and by using a combination of suppression of BC pre-distributed content and user clustering, The utilization rate of the pre-distributed content can be increased.

In the above description, the user set is clustered based on the viewing content of each user on each day, but in reality, the viewing content of each user on the nth day cannot be known in advance at the start time of the nth day. Therefore, it is necessary to perform clustering at the start time of the nth day using information of contents viewed by each user by the (n-1) th day. Therefore, a period Tc for determining x _{i, j} using the past P days in clustering for an arbitrarily defined P is used. That, x _i with respect to the from n-P-th day the n-1_Nichimeno P days x _i for the content j to user i is _watched, set _j = 1, the contents j not watched _{, j} = 0.

12 (a) and 12 (b), when C = 2 and P = 10 and P = ∞, the number of users classified into each cluster generated on each day and the affiliation cluster change on each day A time series of ratios of all users to all users is shown. However, when P = ∞, x _{i, j} was set for all contents viewed from the first day to the nth day. When all the past viewing content histories are taken into account (P = ∞), the amount of change in the viewing content vector x _i of each user i on each day becomes small, so that the number of users whose belonging cluster changes on each day is 0 Less than about 01%. Also, considering only the viewing content history for the most recent 10 days (P = 10), the number of cluster change users is less than about 0.1% of the total, and regardless of the value of P, it is The number of users making up the cluster does not change significantly. About 97% or more of the users are classified into cluster 1.

  FIG. 12C shows a time series of average values of TDLTs for each day of users classified into each cluster. Again, this characteristic does not change greatly throughout the period, and the TDLT of the user classified into the cluster 2 is about 100 times the TDLT of the user of the cluster 1. From these results, it can be confirmed that cluster 2 is composed of a small number of heavy users, whereas cluster 1 is composed of a large number of general users.

  FIG. 12D shows a time series of correlation coefficients between the TDLTs for the contents m of the user set constituting each cluster and the TDLTs for the contents m of all users on each day. However, since there was no significant difference depending on the value of P, only the result when P = 10 is shown. Both clusters have a large correlation coefficient of 0.95 or more, and the TDLT for each content of the user set constituting both clusters shows the same tendency as the TDLT for each content of all users. Therefore, in the case of C = 2, it can be said that the user set is classified by the size of the total viewing amount, not the difference in the tendency of the viewing content.

In FIG. 13A, the centroids mc _{, j} of each cluster c configured on the 100th day when P = 10 are plotted in descending order of the TDLT for each content j of all users on the 100th day. The figure is shown. m _{c, j} corresponds to the proportion of users who have viewed content j at least once in the past P days among the users classified into cluster c, that is, the audience rating. In both clusters, the audience rating of each content is the popularity ranking. It is decreasing with the decline. However, cluster 2 has a higher audience rating for all content than cluster 1. Therefore, in the case of C = 2, it can be confirmed that the cluster is configured by the magnitude of the total viewing amount, not the difference in the tendency of the viewing content. The same was confirmed when C was 4 and 8. However, FIG. 13B shows a graph in which the audience rating of each content on the 100th day is plotted against four of the 16 clusters when C = 16. On the other hand, it can be confirmed that a cluster having a high audience rating of specific content of low popularity has appeared. As described above, when the number of clusters C is increased, it is possible to confirm the tendency that the user set is clustered due to the difference in the tendency of the viewing content.

  FIG. 14 is a diagram in which the number of constituent users of each cluster generated on the 100th day when C is set to 8 or 16 is plotted for the four cases of P in descending order. In any case of P, the number of constituent users between clusters decreases exponentially, and the difference in the number of users between clusters is large. However, the disparity in the number of constituent users decreases as the target period length P of the viewing content history used for clustering increases. When viewing each user's content viewing history in a shorter time, the difference in viewing history between users is emphasized, resulting in a cluster of homogenous users with more viewing tendencies and fewer viewing trends differing from the overall It can be said that there is a tendency to form peculiar clusters.

First, clustering users at the start of each day n, preferentially in descending order of TDLT of each content m of the n-th day of the user set Xc constituting each cluster c generated, the user of each cluster X _c The case where BC pre-delivery is performed is evaluated. This means that an ideal BC pre-delivery content selection is performed for each generated cluster. FIG. 15A shows a graph in which the cache hit rate ξ for all users is plotted against the number of clusters C for the four cases of P. Compared with the result of the ideal clustering shown in FIG. 11A, ξ greatly decreases. The decrease in ξ increases with an increase in C. For example, in the case of C = 16, ξ = 0.360 at ideal clustering, but when clustering is performed in a realistic manner, ξ = 0.328. (When P = 10). However, in this case as well, when ideal BC pre-delivery content selection is performed, ξ improves as C increases, and the effect of clustering users can be seen. Although the effect of P on performance is small, ξ improves slightly as P decreases.

Actually, since the TDLT of each user on the nth day cannot be known at the start time of the nth day, it is necessary to select the BC pre-delivery content using the TDLT of each user on the previous day. That is, the user is the sum of the total time that view the n-1_Nichinioitekakukontentsu m included in the user set X _c constituting each cluster c generated at the beginning of each day n

The BC pre-delivery contents for the users constituting each cluster c are selected in descending order. However, Du _{, m, n-1} is the total time that the user u viewed the content m on the ( _n-1) th day.

  FIG. 15 (b) shows four cases of P.

Is a graph in which the cache hit rate ξ is plotted against C when BC pre-delivery scheduling is performed on day n. When C = 2, ξ improves slightly, but when C becomes larger than about 4, ξ decreases as C increases. It can also be confirmed that the smaller the P is, the larger the degree of decrease in ξ is. To find out these reasons,

FIG. 16 is a diagram in which φ (x) is plotted against x for the six cases of C when P is 10 or ∞. However, s ^c (m, n) is the m-th content ID in descending order of the TDLT of the users constituting the cluster c on the n-th day. φ (x) is the ratio of the sum of the TDLTs of the next day of the top x contents with the largest TDLT in each day of the user set constituting each cluster to the sum of the TDLTs of the top x contents of the next day's TDLT. φ (x) takes a value in a range of 0 <φ (x) ≦ 1, and the larger this value is, the TDLT of the previous day.

This means that the estimation accuracy is high when the TDLT for the content m in the cluster c on the nth day is estimated using.

  As the number of clusters C increases, as shown in FIG. 14, the number of users constituting each cluster decreases. As a result, φ (x) decreases in a wide region of x, and the TDLT of the next day is calculated using the TDLT of the previous day. It can be confirmed that the estimation accuracy in the case of estimation deteriorates. Further, as seen in FIG. 14, the smaller the P, the greater the difference in the number of constituent users in each cluster, and in particular the number of constituent users in a small cluster. Therefore, in the case of P = 10, the degree of decrease in φ (x) is large particularly in the region where x is small compared to the case of P = ∞. For this reason, the estimation accuracy of TDLT for highly popular content is particularly deteriorated when P is small. As a result, as seen in FIG. 15B, when C becomes larger than about 4, ξ decreases as C increases, and as P decreases, the degree of decrease of ξ increases.

  In FIG. 11 (b), P = ∞ is set, and K, V, ξ, and η when BC pre-delivery is performed using the previous day's TDLT are six values of C (2, 3, 4, 8, 16). , 32). Since the deterioration of the accuracy of the viewing content information of each user used for clustering and the deterioration of the estimation accuracy of the TDLT for each content of each cluster increase as C increases, the ideal case shown in FIG. In comparison, as C increases, all the performances deteriorate. In the case of C = 2, the performance is slightly improved as compared with the case of C = 1, but when C is set to 3 or more, the performance deteriorates conversely as C increases.

  As described above, when performing BC pre-distribution, clustering of users based on content viewing history itself is effective, and although the load on the distribution server is reduced as the number of clusters C increases, If the previous day's TDLT is used when selecting the content to be pre-distributed, the error from the next day's TDLT increases as C increases, so clustering is effective only when the number of clusters is reduced to about C = 2 It becomes.

  As described above, according to the present embodiment, the load reduction effect of the distribution server is reduced by performing BC pre-delivery of only a small number of highly popular contents instead of pre-delivering the maximum number of contents. The amount of storage required for the STB can be significantly reduced, and users with similar content viewing trends are clustered, and BC pre-distribution is performed for each cluster, thereby improving the cache hit rate and the distribution server. It is possible to improve the load reduction effect.

  In the above-described embodiment, BC pre-distribution (distribution suppression) of only a few popular contents is combined with clustering (user clustering) of users who are close to viewing contents, but only distribution suppression is performed. You may make it perform. Even if only distribution suppression is performed, the content that is pre-distributed is a small number and highly popular content, so that the cache hit rate is increased. Also, the amount of storage required for the STB is reduced.

  The content pre-distribution method, system, and apparatus of the present invention have a small number of high popularity according to a predetermined distribution schedule regardless of the distribution request, in addition to the distribution form for immediate unicast distribution in response to the user's distribution request. Content pre-distribution method, system, and apparatus for reducing peak distribution server load and network load by distributing only content of all broadcasts to all users and pre-distributing to STBs installed in user premises As described above, it can be used for a VoD service that distributes high-quality video content.

  DESCRIPTION OF SYMBOLS 101 ... Distribution server, 102 ... User side apparatus, 103 ... User cluster design part, 104 ... Delivery schedule design part, 105 ... NW (network), 106 ... Content pre-delivery apparatus.

Claims (9)

Periodically, a content pre-delivery method that pre-distributes popular content by broadcast to all users from the delivery server via the network.
A content pre-distribution method comprising a pre-distribution step of pre-distributing only a small number of highly popular contents than the maximum number of contents that can be pre-distributed during one period of the pre-distribution. Periodically, a content pre-delivery method that pre-distributes popular content by broadcast to all users from the delivery server via the network.
A clustering step of classifying the user into a plurality of clusters having similar viewing tendencies for each period of the pre-delivery;
A pre-delivery step of pre-delivering only a few popular contents less than the maximum number of contents that can be pre-delivered during one period of the pre-delivery for each of the classified clusters. Content pre-delivery method. In the content pre-delivery method according to claim 2,
The clustering step includes
Each of the clustering targets is a binary variable that takes a first value if the user views the content set for clustering within a certain period, and takes a second value if the user does not view the content. A content pre-distribution method characterized by classifying users into a plurality of clusters arbitrarily determined using an average method. In the content pre-delivery method according to claim 2,
The clustering step includes
A content pre-distribution method, comprising: calculating a distance between each user and the center of gravity of each cluster based on a set of contents viewed by the user within a predetermined period and a complement thereof. In the content pre-delivery method according to claim 2,
The pre-delivery step includes
Using the total time that the user viewed the content on the (n-1) th day, the user included in the user set that constitutes each cluster generated at the start time of each nth day, A content pre-distribution method characterized by selecting content to be pre-distributed to the users constituting each cluster in descending order of the total time of viewing. Periodically, in a content pre-distribution system that pre-distributes popular content by broadcast to all users from the distribution server via the network,
A content pre-distribution system comprising pre-distribution means for pre-distributing only a small number of highly popular contents than the maximum number of contents that can be pre-distributed during one period of the pre-distribution. Periodically, in a content pre-distribution system that pre-distributes popular content by broadcast to all users from the distribution server via the network,
Clustering means for classifying a user into a plurality of clusters having a tendency to view content for each period of the pre-delivery;
Pre-distribution means for pre-distributing only a small number of highly popular contents than the maximum number of contents that can be pre-distributed during one period of the pre-distribution for each of the classified clusters, Content pre-delivery system. Periodically, in a content pre-distribution device that pre-distributes popular content by broadcast to all users from the distribution server via the network,
A content pre-distribution device comprising pre-distribution means for pre-distributing only a small number of highly popular contents than the maximum number of contents that can be pre-distributed during one period of the pre-distribution. Periodically, in a content pre-distribution device that pre-distributes popular content by broadcast to all users from the distribution server via the network,
Clustering means for classifying a user into a plurality of clusters having a tendency to view content for each period of the pre-delivery;
Pre-delivery means for pre-delivering only a few popular contents less than the maximum number of contents that can be pre-delivered during one period of the pre-delivery for each of the classified clusters. Content pre-delivery device.

Images