CN110889063B - Video pre-caching method based on Hox process and matrix decomposition - Google Patents

Abstract

The invention provides a video pre-caching method based on a Hox process and matrix decomposition, which comprises the following steps: s1, using hit rate to represent the proportion of cached requests in all requests within a period of time, and re-mathematically defining the hit rate in a mode suitable for a point process according to a hit rate model; s2, predicting the intensity of all videos and sequencing according to the self history of the equipment by using a self-excited Hox process cache model, and caching videos with a certain size; s3, adding the influence of neighbor historical clicks on equipment by the cache model in the process of exciting the Hoxwell, predicting the intensity of all videos, and improving the hit rate; s4, decomposing and reducing the dimension of the parameters of the S3 by using a matrix decomposition dimension reduction model; s5, optimizing the cache model of the process of the mutual excitation Hoxwell by using a process optimization algorithm, and predicting the future behavior intensity. The method and the system can effectively predict the watching intensity of each device to different videos in the future, and can remarkably reduce the flow in the future network and improve the user experience by pre-caching some videos with the maximum intensity.

Description

Video pre-caching method based on Hox process and matrix decomposition

Technical Field

The invention relates to the field of network video caching, in particular to a video pre-caching method based on a Hox process and matrix decomposition.

Background

With the development of online video playing markets, online video playing is becoming more popular, and the number of people watching online videos is also increasing. Cisco predicts that in the future, most of the internet traffic will be video traffic, and on the mobile side, the proportion of traffic generated by the online video service that occupies internet traffic will increase from 59% in 2017 to 79% in 2022. While at the same time the number of online videos is growing at a staggering rate, it is counted that 300 hours of video content are uploaded to YouTube each day, including UGC, news, television shows, movies, etc. Accordingly, there are more and more transmission or terminal devices in the internet, so in order to alleviate the heavy traffic load caused by online video, video content may be cached in various internet devices, including devices such as an edge server, a set-top box, and a personal computer.

First, the simplest way to cache video is to cache the most frequently accessed video, which may also be referred to as the most popular video. However, in the real world, how to select the most popular video faces at least two challenges: first, the user's interests may be affected by various factors such as recommendation, viewer gender, age, etc., and thus the user's video requests may change over time, and the video requested by the user is only a small portion of the most popular video. Meanwhile, because the cache space of each device is limited, devices in the internet need to discard useless videos in time and continuously update the self-cached videos, so that the latest interests of users can be captured in time and higher cache efficiency is achieved. Second, in the internet, the storage and bandwidth capacities of devices are heterogeneous, and the services they provide are also diversified, and thus, it is not possible to simply copy the cache of a certain device to all other devices.

In recent years, video caching methods for predicting future popularity of video by using video history request records are relatively many. Advantageously, the historical popularity of a video can be obtained simply by calculating how frequently the video is requested. However, predicting video popularity is dependent on a large number of user request recordings, i.e., we cannot accurately predict popularity of video on devices serving only one or two users. Thus, popularity-based caching algorithms are useless for edge devices, which cannot capture diverse user interests and diverse videos. The click action of the user can be captured by the point process model, the point process model is expressed in a mathematical form, future click actions of the user can be predicted through the self-excited Hox process, a prediction list can be enriched through the mutual-excited Hox process, and the prediction is more accurate. By utilizing the Hoxwell process, future video watching actions of each device can be effectively predicted, so that the effects of caching in advance and reducing network traffic are achieved.

Disclosure of Invention

In order to solve the defect that video popularity on equipment of a plurality of user services cannot be accurately predicted by utilizing video history request records in the prior art, the invention provides a video pre-caching method based on a Hox process and matrix decomposition.

In order to solve the technical problems, the technical scheme of the invention is as follows:

a video pre-caching method based on a Hox process and matrix decomposition comprises the following steps:

s1, using hit rate to represent the proportion of cached requests in all requests within a period of time, and re-mathematically defining the hit rate in a mode suitable for a point process according to a hit rate model;

s2, predicting the intensity of all videos and sequencing according to the self history of the equipment by using a self-excited Hox process cache model, and caching videos with a certain size;

s3, adding the influence of neighbor historical clicks on equipment by the cache model in the process of exciting the Hoxwell, predicting the intensity of all videos, and improving the hit rate;

s4, decomposing and reducing the dimension of the parameters of the S3 by using a matrix decomposition dimension reduction model;

s5, optimizing the cache model of the process of the mutual excitation Hoxwell by using a process optimization algorithm, and predicting the future behavior intensity.

In a preferred embodiment, S1 includes the following steps:

s11, using

To represent the intensity under the hoxwell process, where u represents the buffering device, i represents the video, and t represents the time at which the intensity is located;

s12, defining the hit rate as follows:

wherein s is _i Is the size of the buffered video, B _u For the buffer space of device u +.>

Is an indication function, indicating whether device u will cache video i at time t,

indication function

Is defined as follows:

s13, the hit rate model, namely the objective function, is expressed as:

in a preferred embodiment, the specific steps at S2 are as follows:

s21, definition epsilon ^T ＝{t ₁ ,t ₂ ,…,t _K -indicating the click times of all click events, the recorded time window being]0, T), and has t ₁ ＜t ₂ ＜…＜t _K K represents the total number of events;

s22 definition

Then

Wherein b _ui Representing the reference or immigrating strength of the current process, i.e. the strength value when no history event has occurred in the process, phi (t-t') represents the activation function of the inter-excited hox process, i.e.:

φ _ui (t-t′)＝α _ui g(t-t′)#(2-4)

α _ui >0 indicates how much the current process will be affected by the history of events, and for g (t-t'), there is:

g(t-t′)＝exp(-δ(t-t′))#(2-5)

delta >0 is a hyper-parameter representing the decay rate of the activation function or decay function, the greater the value, the faster the decay;

s23, optimizing each process by using an optimization method of the Hox process in order to maximize the hit rate, wherein a likelihood function is defined as follows:

after likelihood functions are found, the likelihood functions are derived, and then the gradient descent method is used for estimating each parameter.

In a preferred embodiment, the specific step of S3 is as follows:

s31, adding influence of neighbor history electrolysis on current equipment, wherein the definition is as follows:

s32, combining (2-3) with (2-7) to obtain a historical click event sequence which considers both self historical click event sequences and historical click sequences of other devices

Is calculated according to the estimated formula:

wherein delta is used ₁ And delta ₂ The different attenuation levels of the two attenuation functions are respectively represented and defined, so the attenuation functions are represented as follows:

g ₁ (t-t′)＝exp(-δ ₁ (t-t′))#(2-9)

g ₂ (t-t′)＝exp(-δ ₂ (t-t′))#(2-10)

s33 definition of SE and ME

Each device receives the self-clicking process and other devicesThe impact of the standby click process is different, so that one parameter, beta epsilon (0, 1), needs to be added to balance SE and ME, and then,

the expression is as follows:

in a preferred embodiment, the specific step of S4 is as follows:

s41, performing dimension reduction on the formula (2-13), referring to SVD algorithm, and performing dimension reduction on b _ui And alpha _ui The following approximations are respectively made:

b _ui ＝b _u +b _i #(2-14)

wherein b _u ,b _i Representing the fundamental intensity (intensity of migration) of the device and video, respectively, for alpha _ui Matrix decomposition is carried out, q _i ,p _u The video implicit vectors respectively representing alpha and the implicit vectors of the device are defined as d, so that the parameter quantity is reduced from 2mn to n+m+nd+md, and m, n > d.

S42, defining other device sets affecting the device u as R _u Improvements in or relating to

ME part of (3), finally->

The estimated expression of (2) is:

so far, the fitting and generalization capability is stronger, and the expression meaning is wider

Formula (2-16).

In a preferred embodiment, S5 includes the following steps;

s51, solving parameters according to the estimation formula obtained in the step S4, and maximizing an equivalent log likelihood function:

wherein, formula (2-17) is a log-likelihood function of the click process of a single device-video, then for all devices and videos, there is a log-likelihood function as follows:

maximizing equation (2-18), equivalent to minimizing-L, to prevent overfitting, increasing generalization ability, for b _u ,b _i ,q _i ,p _u And respectively adding quadratic regularization terms to obtain a final optimization target:

s.t.b _u ,b _i ,q _i ,p _u >0,

β _u ∈(0,1),

s52, requiring all the parameters in the formulas (2-19) to be greater than 0 to prevent

The logarithmic calculation is illegal, the method is to be used2-16) after substitution (2-18), to obtain:

s53, performing bias guide on the components (2-20) to obtain a formula of each parameter under gradient update:

/>

s54, solving by using a gradient descent method according to a gradient formula of each parameter, and assuming that one parameter is theta and the initial value is theta ₀ Then the update formula is:

according to specific scenes or data, different initial values are set, and then a gradient descent method is used for obtaining sub-optimal solutions of all parameters.

Compared with the prior art, the technical scheme of the invention has the beneficial effects that:

the method and the system can effectively predict the watching intensity of each device to different videos in the future, and can remarkably reduce the flow in the future network and improve the user experience by pre-caching some videos with the maximum intensity.

Drawings

FIG. 1 is a flow chart of the steps of the present invention.

Detailed Description

The drawings are for illustrative purposes only and are not to be construed as limiting the present patent;

for the purpose of better illustrating the embodiments, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the actual product dimensions;

it will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.

The technical scheme of the invention is further described below with reference to the accompanying drawings and examples.

Example 1

As shown in fig. 1, a video pre-caching method based on a hough process and matrix decomposition includes the following steps:

s1, using hit rate to represent the proportion of cached requests in all requests within a period of time, and re-mathematically defining the hit rate in a mode suitable for a point process according to a hit rate model;

s2, predicting the intensity of all videos and sequencing according to the self history of the equipment by using a self-excited Hox process cache model, and caching videos with a certain size;

s3, adding the influence of neighbor historical clicks on equipment by the cache model in the process of exciting the Hoxwell, predicting the intensity of all videos, and improving the hit rate;

s4, decomposing and reducing the dimension of the parameters of the S3 by using a matrix decomposition dimension reduction model;

s5, optimizing the cache model of the process of the mutual excitation Hoxwell by using a process optimization algorithm, and predicting the future behavior intensity.

Example 2

The video pre-caching method based on the hox process and matrix decomposition provided in this embodiment is consistent with the above provided method, and only the steps are further limited.

A video pre-caching method based on a Hox process and matrix decomposition comprises the following steps:

s1, using hit rate to represent the proportion of cached requests in all requests within a period of time, and re-mathematically defining the hit rate in a mode suitable for a point process according to a hit rate model;

s2, predicting the intensity of all videos and sequencing according to the self history of the equipment by using a self-excited Hox process cache model, and caching videos with a certain size;

s3, adding the influence of neighbor historical clicks on equipment by the cache model in the process of exciting the Hoxwell, predicting the intensity of all videos, and improving the hit rate;

s4, decomposing and reducing the dimension of the parameters of the S3 by using a matrix decomposition dimension reduction model;

s5, optimizing the cache model of the process of the mutual excitation Hoxwell by using a process optimization algorithm, and predicting the future behavior intensity.

In a preferred embodiment, S1 includes the following steps:

s11, using

To represent the intensity under the hoxwell process, where u represents the buffering device, i represents the video, and t represents the time at which the intensity is located;

s12, defining the hit rate as follows:

wherein s is _i Is the size of the buffered video, B _u For the buffer space of device u +.>

Is an indication function, indicating whether device u will cache video i at time t,

indication function

Is defined as follows:

s13, the hit rate model, namely the objective function, is expressed as:

/>

in a preferred embodiment, the specific steps at S2 are as follows:

s21, definition epsilon ^T ＝{t ₁ ,t ₂ ,…,t _K -indicating the click times of all click events, the time window recorded is [0, t), and there is t ₁ ＜t ₂ ＜…＜t _K K represents the total number of events;

s22 definition

Then

Wherein b _ui Representing the reference or immigrating strength of the current process, i.e. the strength value when no history event has occurred in the process, phi (t-t') represents the activation function of the inter-excited hox process, i.e.:

φ _ui (t-t′)＝α _ui g(t-t′)#(2-4)

α _ui >0 indicates how much the current process will be affected by the history of events, and for g (t-t'), there is:

g(t-t′)＝exp(-δ(t-t′))#(2-5)

delta >0 is a hyper-parameter representing the decay rate of the activation function or decay function, the greater the value, the faster the decay;

s23, optimizing each process by using an optimization method of the Hox process in order to maximize the hit rate, wherein a likelihood function is defined as follows:

after likelihood functions are found, the likelihood functions are derived, and then the gradient descent method is used for estimating each parameter.

In a preferred embodiment, the specific step of S3 is as follows:

s31, adding influence of neighbor history electrolysis on current equipment, wherein the definition is as follows:

s32, combining (2-3) with (2-7) to obtain a historical click event sequence which considers both self historical click event sequences and historical click sequences of other devices

Is calculated according to the estimated formula:

wherein delta is used ₁ And delta ₂ The different attenuation levels of the two attenuation functions are respectively represented and defined, so the attenuation functions are represented as follows:

g ₁ (t-t′)＝exp(-δ ₁ (t-t′))#(2-9)

g ₂ (t-t′)＝exp(-δ ₂ (t-t′))#(2-10)

s33 definition of SE and ME

/>

The impact of each device receiving its own click process and the other device click process is different, so a parameter βe (0, 1) needs to be added to balance SE and ME, and,

the expression is as follows:

in a preferred embodiment, the specific step of S4 is as follows:

s41, performing dimension reduction on the formula (2-13), referring to SVD algorithm, and performing dimension reduction on b _ui And alpha _ui The following approximations are respectively made:

b _ui ＝b _u +b _i #(2-14)

wherein b _u ,b _i Representing the fundamental intensity (intensity of migration) of the device and video, respectively, for alpha _ui Matrix decomposition is carried out, q _i ,p _u The video implicit vectors respectively representing alpha and the implicit vectors of the device are defined as d, so that the parameter quantity is reduced from 2mn to n+m+nd+md, and m, n > d.

S42, defining other device sets affecting the device u as R _u Improvements in or relating to

ME part of (3), finally->

The estimated expression of (2) is:

to this end, a fitting sum is obtainedHas stronger generalization ability and wider expression meaning

Formula (2-16).

In a preferred embodiment, S5 includes the following steps;

s51, solving parameters according to the estimation formula obtained in the step S4, and maximizing an equivalent log likelihood function:

wherein, formula (2-17) is a log-likelihood function of the click process of a single device-video, then for all devices and videos, there is a log-likelihood function as follows:

maximizing equation (2-18), equivalent to minimizing-L, to prevent overfitting, increasing generalization ability, for b _u ,b _i ,q _i ,p _u And respectively adding quadratic regularization terms to obtain a final optimization target:

s.t.b _u ,b _i ,q _i ,p _u >0,

β _u ∈(0,1),

/>

s52, requiring all the parameters in the formulas (2-19) to be greater than 0 to prevent

Resulting in illegal logarithmic calculation, substituting (2-16) into (2-18) yields:

s53, performing bias guide on the components (2-20) to obtain a formula of each parameter under gradient update:

s54, solving by using a gradient descent method according to a gradient formula of each parameter, and assuming that one parameter is theta and the initial value is theta ₀ Then the update formula is:

according to specific scenes or data, different initial values are set, and then a gradient descent method is used for obtaining sub-optimal solutions of all parameters.

The terms describing the positional relationship in the drawings are merely illustrative, and are not to be construed as limiting the present patent;

it is to be understood that the above examples of the present invention are provided by way of illustration only and not by way of limitation of the embodiments of the present invention. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary here nor is it exhaustive of all embodiments. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are desired to be protected by the following claims.

Claims (4)

1. The video pre-caching method based on the Hox process and matrix decomposition is characterized by comprising the following steps of:

s1, using hit rate to represent the proportion of cached requests in all requests within a period of time, and according to a hit rate model, using a mode suitable for a click process to mathematically define the hit rate again; wherein:

s11, using

To represent the intensity under the hoxwell process, where u represents the buffering device, i represents the video, and t represents the time at which the intensity is located;

s12, defining the hit rate as follows:

wherein s is _i Is the size of the buffered video, B _u For the buffer space of device u +.>

Is an indication function, indicating whether device u will cache video i at time t,

indication function

Is defined as follows:

s13, the hit rate model, namely the objective function, is expressed as:

s2, predicting the intensity of all videos and sequencing according to the self history of the equipment by using a self-excited Hox process cache model, and caching videos with a certain size;

s3, adding the influence of neighbor historical clicks on equipment by the cache model in the process of exciting the Hoxwell, predicting the intensity of all videos, and improving the hit rate;

s4, decomposing and reducing the dimension of the parameters of the S3 by using a matrix decomposition dimension reduction model;

s5, optimizing the cache model of the process of the mutual excitation Hoxwell by using a process optimization algorithm, and predicting the future behavior intensity;

the specific steps at S2 are as follows:

s21, definition epsilon ^T ＝{t ₁ ,t ₂ ,…,t _K -indicating the click times of all click events, the time window recorded is [0, t), and there is t ₁ <t ₂ <…<t _K K represents the total number of events;

s22 definition

Then

Wherein b _ui Representing the reference or immigrating strength of the current process, i.e. the strength value, phi, when no historical events occur in the process _ui (t-t') represents the activation function of the inter-excited hox process, namely:

φ _ui (t-t′)＝α _ui g(t-t′) (2-4)

α _ui >0 indicates how much the current process will be affected by the historical events, and for g (t-t'), there is:

g(t-t′)＝exp(-δ(t-t′)) (2-5)

delta >0 is a hyper-parameter representing the decay rate of the activation function or decay function, the greater the value, the faster the decay;

s23, optimizing each process by using an optimization method of the Hox process in order to maximize the hit rate, wherein a likelihood function is defined as follows:

after likelihood functions are found, the likelihood functions are derived, and then the gradient descent method is used for estimating each parameter.

2. The video pre-caching method based on the hough process and the matrix decomposition according to claim 1, wherein the specific step of S3 is as follows:

s31, adding influence of neighbor history clicking on current equipment, wherein the definition is as follows:

s32, combining (2-3) with (2-7) to obtain a historical click event sequence which considers both self historical click event sequences and historical click sequences of other devices

Is calculated according to the estimated formula:

wherein delta is used ₁ And delta ₂ The different attenuation levels of the two attenuation functions are respectively represented and defined, so the attenuation functions are represented as follows:

g ₁ (t-t′)＝exp(-δ ₁ (t-t′)) (2-9)

g ₂ (t-t′)＝exp(-δ ₂ (t-t′)) (2-10)

s33 definition of SE and ME

Each device is affected differently by the clicking process of itself and the clicking process of the other device, so that a weight adjustment parameter beta needs to be added _u E (0, 1) to trade-off the effects of SE and ME, then,

the expression is as follows:

3. the video pre-caching method based on the hough process and the matrix decomposition according to claim 2, wherein the specific step of S4 is as follows:

s41, performing dimension reduction on the formula (2-13), referring to SVD algorithm, and performing dimension reduction on b _ui And alpha _ui The following approximations are respectively made:

b _ui ＝b _u +b _i (2-14)

wherein b _u ，b _i Representing the fundamental intensity (intensity of migration) of the device and video, respectively, for alpha _ui Matrix decomposition is carried out, q _i ，p _u The video implicit vector of alpha is respectively represented, the implicit vector of the device is defined as d, so that the parameter quantity is reduced from 2mn to n+m+nd+md, and m, n > d;

s42, defining other device sets affecting the device u as R _u Improvements in or relating to

ME part of (3), finally->

The estimated expression of (2) is:

so far, the fitting and generalization capability is stronger, and the expression meaning is wider

Formula (2-16).

4. A video pre-buffering method based on a hough process and matrix decomposition according to claim 3, wherein said S5 comprises the steps of;

s51, solving parameters according to the estimation formula obtained in the step S4, and maximizing an equivalent log likelihood function:

wherein, formula (2-17) is a log-likelihood function of the click process of a single device-video, then for all devices and videos, there is a log-likelihood function as follows:

maximizing equation (2-18), equivalent to minimizing-L, to prevent overfitting, increasing generalization ability, for b _u ,b _i ,q _i ,p _u And respectively adding quadratic regularization terms to obtain a final optimization target:

s52, requiring all the parameters in the formulas (2-19) to be greater than 0 to prevent

Resulting in illegal logarithmic calculation, substituting (2-16) into (2-18) yields:

s53, performing bias guide on the components (2-20) to obtain a formula of each parameter under gradient update:

/>

s54, solving by using a gradient descent method according to a gradient formula of each parameter, and assuming that one parameter is theta and the initial value is theta ₀ Then the update formula is:

according to specific scenes or data, different initial values are set, and then a gradient descent method is used for obtaining sub-optimal solutions of all parameters.

Images