CN104679992A - Markov model designing method based on user service use time - Google Patents

Abstract

The invention discloses a Markov model designing method based on user service use time. The method is characterized in that a value difference metric (VDM) algorithm is used to perform discretization on the user service use time so as to lower calculating complexity, time attribute domains are divided into intervals, interval marks are used to replace actual time values, then the state transfer matrix in a Markov model is adjusted by using the user service access trend weight, and accordingly prediction algorithm accuracy is increased effectively.

Description

Based on the method for designing of the Markov model of customer service service time

Technical field

The present invention relates to a kind of wireless general at ambient based on the method for designing of the Markov model of customer service service time, belong to the technical field of communication.

Background technology

Since entering 21 century, along with the development of microelectronics, chip and ICT (information and communication technology), the mankind will progressively enter the Ubiquitous Network epoch, wherein wireless Ubiquitous Network, that is: wireless ubiquitous environment is that people places oneself in the midst of among immanent network, realize people at any time and place, use the message exchange of any network and anyone and thing, based on individual and social demand, utilize existing network technology and new network technology, for individual and society provide ubiquitous, without the information service do not contained and application.Along with the fast development of Internet technology, service resources also presents the situation of explosive growth thereupon, how from the service resources of magnanimity, obtains the business needed for user quickly and efficiently, make user obtain higher Consumer's Experience, become the major fields of concern.Wireless at ambient general, by customer-centric, the generation of business is directed to the demand of user, multimode terminal can access different networks, how to realize effective utilization of Internet resources between different radio Access Network and terminal capability, improves formation transfer service quality, from different perspectives, different levels are met consumers' demand, improve user experience quality and provide Intelligent Service for user, therefore, being necessary to study user's behavior prediction.

Wireless at ambient general in future, 2G/3G/4G cellular mobile communications networks, IEEE 802.11 WLAN (wireless local area network) (WLAN) and IEEE 802.16 wireless MAN (WiMAX) using as main flow wireless access way and deposit, to be interconnected by wired backbone or Wireless Mesh network etc. between access network, also will realize Ad Hoc by series techniques such as IEEE 802.11 or IEEE 802.15 between user terminal to connect, thus provide ubiquitous wireless access for ubiquitous business demand; On the other hand, in recent years along with mobile intelligent terminal and Mobile operating system fast development, further enhancing the Consumer's Experience of mobile terminal.Almost all at present cable Internet service types can be transplanted on mobile Internet, provide appointed condition for realizing business wireless penetration.

Wireless traffic personalization has become the new focus of current business development, and along with becoming clear day by day of user individual feature, personalized push even develops into the new marketing ideas of operator, service provider and trend.Therefore personalized push technology is arisen at the historic moment, it can from a large number of users behavior historical information, by various mining algorithms and forecast model, automatically the business meeting its demand is pushed out moderately and at the right moment to user, for user provides personalized business customizing service and business recommended, provide decision-making foundation for service provider releases new business simultaneously.

At present, excavation mainly laid particular emphasis on to the research of user behavior both at home and abroad and analyze user behavior characteristic sum rule and user's behavior prediction aspect.The instrument that correlative study adopts comprises data mining algorithm and Markov forecast model etc.Wherein Markov model is simply a kind of and effective model.Due to the markov property of Markov chain, Markov model theory is mostly adopted to predict user behavior at present.Markov model with its large information reserved, thus ensure that it to predict feature accurately and becomes a typical model of this respect research.

But by carrying out the analysis of customer service Forecasting Methodology to existing various utilization Markov model, a very important factor in the construction process of model: the service time of customer service is not taken into account, that is in existing method, only considered the type of customer service, think that each customer service time the same.But different business service time is very important for the interest of performance user, when business service time hour should not consider in probability.The method of these predictions considers that merely type of service often obtains the identical result of multiple probability, just the business of bad unique determination user needs in future, and precision of prediction is not high enough.And the present invention can solve problem above well.

Summary of the invention

The object of the invention is to propose a kind of method for designing of the Markov model based on customer service service time, this method solve customer service service time to the problem predicted the outcome, the method utilizes customer service to access tendency weight and adjusts the state-transition matrix in Markov model, effectively improves the accuracy of prediction.On the one hand, the learning process of Markov model Forecasting Methodology new in the method has good theoretical property, and prediction accuracy is high; On the other hand, the method is simple to operate, is easy to realize, and has good application prospect.

The present invention solves the technical scheme that its technical matters takes: a kind of method for designing of the Markov model based on customer service service time, first the method uses value difference indexing amount algorithm (that is: VDM) to carry out sliding-model control to the service time of customer service, thus reduce the complicacy of calculating, time attribute territory is divided into interval, actual time value is replaced with the label in interval, on this basis, utilize customer service to access tendency weight adjust the state-transition matrix in Markov model, thus effectively improve prediction algorithm ground accuracy.

Method flow:

The invention provides a kind of method for designing of the Markov model based on customer service service time, the method comprises the steps:

Step 1: according to historic user business sample sequence, the ordinary practice of business is used in conjunction with user, think that user is generally divided into short stay, common use, long-time use 3 kinds of modes to the time of the use of business, determine that the target of discretize obtains three discrete segments;

Step 2: use and use value difference indexing amount algorithm (that is: VDM) to carry out sliding-model control service time to customer service;

Step 3: according to customer service sample sequence structure Markov model, thus obtain corresponding state-transition matrix;

Step 4: on the basis of step 3, according to customer service access tendency weight adjusting state-transition matrix, thus constructs the Markov model based on customer service service time;

Step 5: predict according to the future services of above-mentioned Markov model to user.

Value difference indexing amount algorithm (VDM) in the present invention is applied to the discretize of connection attribute, there is good technique effect, discretization technique is applied in the expression of customer service service time, time attribute territory is divided into interval, replace actual time value with the label in interval, thus reduce the complicacy of calculating.

Beneficial effect:

1, the present invention utilizes customer service access tendency weight to adjust the state-transition matrix in Markov model, thus effectively improves the accuracy of prediction algorithm.

2, customer service this factor service time is taken into account by the present invention, adds business this factor service time when constructing Markov model, thus reaches the target improving user's behavior prediction accuracy.

3, the method for designing of the Markov model based on customer service service time that produces of the present invention, it is simple to operate and be easy to realize, and has good application prospect.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention.

Embodiment

Below in conjunction with Figure of description, innovation and creation are described in further detail.

Embodiment one

As shown in Figure 1, the present invention proposes a kind of method for designing of the Markov model based on customer service service time, the method has merged value difference indexing amount algorithm (that is: VDM) and Markov model, by joining in the construction process of Markov model to the sliding-model control of customer service service time by this factor of time, effectively improve the accuracy of user's behavior prediction.

The present invention mainly comprises two aspects: use value difference indexing amount algorithm (that is: VDM) to carry out sliding-model control service time to customer service on the one hand, discretization technique is applied in the expression of customer service service time, time attribute territory is divided into interval, replace actual time value with the label in interval, thus reduce the complicacy of calculating; Be the structure of Markov model on the other hand, utilize customer service to access tendency weight and the state-transition matrix in Markov model is adjusted, construct the Markov model made new advances.

1. the discretize of customer service service time

Consider the complicacy of calculating, the present invention proposes discretization method to be applied in the expression of customer service service time, and time attribute territory is divided into interval, replaces actual time value with the label in interval.The discretize that value difference indexing amount algorithm (that is: VDM) is applied to connection attribute also can have goodish technique effect.Generally, the distance definition between Category Attributes is:

$\mathrm{\δ}(v_1,v_2)=\left\{\begin{array}{cc}1& v_1\≠v_2\\ 0& v_1=v_2\end{array}\right.$

If current attribute has t discrete value c ₁, c ₂..., c _t, s classification d ₁, d ₂..., d _s, then shown contingency table can be constructed as follows between attribute and category label.

Wherein n _ijrepresent that aggregate attribute value is c _jclassification is d simultaneously _iexample number.N _i*represent that in set, classification is d _iexample number, n _{* j}represent that attribute value is c _jexample number, n _*for all example sums.

When Category Attributes is with category label, by after the relation of investigating attribute value and classification, the distance between attribute different value can be measured out more accurately.Value difference indexing amount method is:

$\mathrm{\δ}(c_p,c_q)=\underset{i=1}{\overset{s}{\mathrm{\Σ}}}|\frac{n_{\mathrm{ip}}}{n_{*p}}-\frac{n_{\mathrm{ip}}}{n_{*p}}|\×\frac{n_{i*}}{n_{**}}$

C in formula _pand c _qrepresent two values of attribute.Intuitively, if c _pand c _qdistribution on of all categories is consistent, then can think that the difference of two attributes is 0.If it is more inconsistent to distribute, then its difference is larger.The span of obvious above-mentioned tolerance is 0≤δ≤1.When distributing consistent, difference is 0; When attribute and classification mutually each other sufficient and necessary condition time, difference is 1, now maximum.

Therefore, the criterion when distance between the discrete value utilizing value difference indexing amount algorithm to calculate can be used as merger or divide connection attribute, for merger, the method comprises as follows:

Input: one group of certain connection attribute different, ascending order is arranged as x ₁, x ₂..., x _nvalue, and the category label of their correspondences; Setting maximum attribute discrete value number V _max, can merger minimum degree value M _min.Parameter V _maxthat discretize expects the result reached, M _minfor regulating discrete effect.

Export: some adjacent intervals, each interval represents a discrete value.

Step 1: use x ₁, x ₂..., x _nform n interval, each interval represents a discrete value, and wherein min represents the lower limit of connection attribute, and max represents the possible maximum value of attribute.

c ₁＝[min,(x ₁+x ₂)/2]

c ₂＝[(x ₁+x ₂)/2,(x ₂+x ₃)/2]

…

c _n＝[(x _n-1+x _n)/2,max]

Step 2: calculate the best adjacent interval with minimum degree value

Step 3: merge best adjacent interval

If (m _minbe less than M _minor k is greater than V _maxif)/* zone distance can merge or discrete segment number is more than the merger result estimated, then combine interval */Then

The present invention uses value difference indexing amount algorithm (that is: VDM) to carry out sliding-model control service time to customer service, value corresponding to parameters in VDM algorithm, is considered as the c in VDM algorithm the service time (in seconds) of every bar business record _j, and the visitor IP (representing different users) of every bar record is considered as the d in VDM algorithm _i.According to historic survey achievement, the ordinary practice of business is used in conjunction with user, think that user is generally divided into short stay, common use, long-time use 3 kinds to the time of the use of business, so the target of discretize obtains three discrete segments, so the V in VDM algorithm _maxbe decided to be 3.Use the discretize that value difference indexing amount algorithm just can realize business use, and three interval intersection values can be obtained be respectively T ₁second, T ₂second.Browsing time is T ₁below second is short stay, T ₂second above be long-time use, between therebetween be common use, discretize result is as shown in the table.

Discrete value	Business time service time (s represents second)
		1	0s<time<＝T 1s
2	T 1s<time<＝T 2s
		3	time>T 2s

The structure of 2.Markov model

Markov model is made up of three parts, comprising:

(1) one group of state Status.Corresponding to the set of all business, be denoted as Status={S, s ₁, s ₂..., s _n, F}.S, F are two virtual states, correspond to and start and final state.

(2) state-transition matrix A.Corresponding element is p _ij, be defined as P (s _j| s _i), namely user is from current business s _itransfer to business s _jprobability.

$A=\left(p_{\mathrm{ij}}\right)=\left[\begin{array}{cccc}{p}_{11}& p_{12}& ...& p_{1n}\\ p_{21}& p_{22}& ...& p_{2n}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ p_{n1}& p_{n2}& ...& p_{\mathrm{nn}}\end{array}\right]$

(3) one groups of probability.The probability of correspondence and each state.

With vectorial H (t)=(0,0 ..., 1) represent that user is in the state of moment t, if now user is in state s ₁, then the i-th dimension of this vector equals 1, and all the other each dimensions are all 0.The state probability vector in moment t system is represented, the probability of every one-dimensional representation different conditions with vectorial V (t).Then can make prediction in the state of moment t to user according to following formula:

V(t)＝H(t-1)×A

In vectorial V (t), the state corresponding to that one dimension that probable value is maximum is exactly that user is in the most probable state of moment t.

On the basis of customer service discretize service time, utilize customer service to access tendency weight and the state-transition matrix in Markov model is adjusted.Wherein customer service access tendency weight w _icomputing formula be: wherein t ₁, t ₂..., t _nfor the discrete value of customer service service time, n is the type sum of business.The adjustment formula of Markov model transition matrix $p_{\mathrm{ij}}^{\′}=\frac{p_{\mathrm{ij}}\·w_j}{\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{p}_{\mathrm{ik}}\·w_k},0\≤i,j\≤n,n$ For state number.

The Markov model construction process that the present invention is based on customer service service time comprises:

(1) constitutional diagram G is generated according to customer service type sequence intersection: if s _iand s _jbe the connected reference business in a sequence, just in constitutional diagram G, add a s _ito s _jlimit e _i,jif, e _i,jexist, just the counting count value on this limit is added 1;

(2) the constitutional diagram G constructed according to above-mentioned steps 1 generates state-transition matrix A:p _ijvalue be e _i,jcount value than upper s _iall go out the count value sum on limit;

(3) according to customer service access tendency weight adjusting A: according to formula p is recalculated in conjunction with customer service access tendency weight _ijvalue, generate new state-transition matrix.

In sum, the invention provides a kind of method for designing of the Markov model based on customer service service time, the method comprises the steps:

Step 1: according to historic user business sample sequence, the ordinary practice of business is used in conjunction with user, think that user is generally divided into short stay, common use, long-time use 3 kinds to the time of the use of business, determine that the target of discretize obtains three discrete segments;

Step 2: use value difference indexing amount algorithm (that is: VDM) to carry out sliding-model control service time to customer service;

Step 3: according to customer service sample sequence structure Markov model, thus obtain corresponding state-transition matrix;

Step 4: on the basis of step 3, according to customer service access tendency weight adjusting state-transition matrix, thus constructs the Markov model based on customer service service time

Step 5: predict according to the future services of above-mentioned Markov model to user.

Embodiment two

The concrete implementation method of the present invention comprises the steps:

A, consider the complicacy of calculating, the present invention proposes discretization technique to be applied in the expression of customer service service time, and time attribute territory is divided into interval, replaces actual time value with the label in interval.

Criterion when distance between the discrete value that value difference indexing amount algorithm calculates can be used as merger or divide connection attribute, for merger, its algorithm is as follows:

Input: one group of certain connection attribute different, ascending order is arranged as x ₁, x ₂..., x _nvalue, and the category label of their correspondences; Setting maximum attribute discrete value number V _max, can merger minimum degree value M _min.Parameter V _maxthat discretize expects the result reached, M _minfor regulating discrete effect.

Export: some adjacent intervals, each interval represents a discrete value.

Step 1: use x ₁, x ₂..., x _nform n interval, each interval represents a discrete value, and wherein min represents the lower limit of connection attribute, and max represents the possible maximum value of attribute.

c ₁＝[min,(x ₁+x ₂)/2]

c ₂＝[(x ₁+x ₂)/2,(x ₂+x ₃)/2]

…

c _n＝[(x _n-1+x _n)/2,max]

Step 2: calculate the best adjacent interval with minimum degree value

Step 3: merge best adjacent interval

If (m _minbe less than M _minor k is greater than V _maxif) Then/* zone distance can merge or discrete segment number than estimate merger result many, then combine interval */

B, utilization value difference indexing amount algorithm (VDM) carry out sliding-model control to customer service, the value corresponding to the parameters in VDM algorithm service time: be considered as the c in VDM algorithm the service time (in seconds) of every bar business record here _j, and the visitor IP (representing different users) of every bar record is considered as the d in VDM algorithm _i.According to historic survey achievement, the ordinary practice of business is used in conjunction with user, think that user is generally divided into short stay, common use, long-time use 3 kinds to the time of the use of business, so the target of discretize obtains three discrete segments, so the V in VDM algorithm _maxbe decided to be 3.Use the discretize that value difference indexing amount algorithm just can realize business use, and three interval intersection values can be obtained be respectively T ₁second, T ₂second.Browsing time is T ₁below second is short stay, T ₂second above be long-time use, between therebetween be common use, discretize result is as shown in the table.

Discrete value	Business time service time (s represents second)
		1	0s<time<＝T 1s
2	T 1s<time<＝T 2s
		3	time>T 2s

C, on the basis of customer service discretize service time, utilize customer service access tendency weight the state-transition matrix in Markov model is adjusted.Wherein customer service access tendency weight w _icomputing formula be: wherein t ₁, t ₂..., t _nfor the discrete value of customer service service time, n is the type sum of business.The adjustment formula of Markov model transition matrix $p_{\mathrm{ij}}^{\′}=\frac{p_{\mathrm{ij}}\·w_j}{\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{p}_{\mathrm{ik}}\·w_k},0\≤i,j\≤n,n$ For state number.

Above to the detailed introduction of the method for designing of a kind of Markov model based on customer service service time provided by the invention, for those skilled in the art, according to the thought of the embodiment of the present invention, all will change in specific embodiments and applications, in sum, embodiments of the invention should not be construed as limitation of the present invention.

Claims (6)

1., based on a method for designing for the Markov model of customer service service time, it is characterized in that, described method comprises the steps:

Step 1: according to historic user business sample sequence, use the ordinary practice of business in conjunction with user, thinks that user is divided into short stay, common use the time to the use of business, uses for a long time, determines that the target of discretize obtains three discrete segments;

Step 2: use and use value difference indexing amount algorithm to carry out sliding-model control service time to customer service;

Step 3: according to customer service sample sequence structure Markov model, thus obtain corresponding state-transition matrix;

Step 4: on the basis of above-mentioned steps 3, according to customer service access tendency weight adjusting state-transition matrix, thus constructs the Markov model based on customer service service time;

Step 5: predict according to the future services of above-mentioned Markov model to user.

2. the method for designing of a kind of Markov model based on customer service service time according to claim 1, it is characterized in that, described method comprises: use value difference indexing amount algorithm to carry out sliding-model control service time to customer service, discretization technique is applied in the expression of customer service service time, time attribute territory is divided into interval, replaces actual time value with the label in interval.

3. the method for designing of a kind of Markov model based on customer service service time according to claim 1, it is characterized in that, described method is structure Markov model, utilize customer service to access tendency weight to adjust the state-transition matrix in Markov model, construct the Markov model made new advances.

4. the method for designing of a kind of Markov model based on customer service service time according to claim 1 or 3, is characterized in that: the Markov model of described method comprises:

(1) one group of state Status; Corresponding to the set of all business, be denoted as Status={S, s ₁, s ₂..., s _n, F}, S, F are two virtual states, correspond to and start and final state;

(2) state-transition matrix A, corresponding element is p _ij, be defined as P (s _j| s _i), namely user is from current business s _itransfer to business s _jprobability;

$A=\left(p_{\mathrm{ij}}\right)=\left[\begin{array}{cccc}{p}_{11}& p_{12}& ...& p_{1n}\\ p_{21}& p_{22}& ...& p_{2n}\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ p_{n1}& p_{n2}& ...& p_{\mathrm{nn}}\end{array}\right]$

(3) one groups of probability, the probability of correspondence and each state.

5. the method for designing of a kind of Markov model based on customer service service time according to claim 1 or 3, is characterized in that: described method is according to customer service access tendency weight w _icomputing formula: wherein t ₁, t ₂..., t _nfor the discrete value of customer service service time, n is the type sum of business, then uses the adjustment formula of Markov model transition matrix n is that state number builds the Markov model made new advances.

6. the method for designing of a kind of Markov model based on customer service service time according to claim 1, it is characterized in that, the value difference indexing amount algorithm application of described method is in the discretize of connection attribute; Discretization technique is applied in the expression of customer service service time; Time attribute territory is divided into interval, replaces actual time value with the label in interval.