Fuzzy Rules-based Classi?cation for Churn
Prediction
Bingquan Huang,Ying Huang,and M-T.Kechadi
School of Computer Science and Informatics,
University College Dublin,
Bel?eld,Dublin4,Ireland
bingquan.huang@ucd.ie
ying.huang.1@ucdconnect.ie
Abstract.The customer churn in?uences the revenue and becomes a
serious problem for a service company.In order to retain the valuable
customers,rule-based classi?cation methods which can transparently in-
terpret the prediction results,have been prefered to use for the analysis
of customer behaviour.However,due to the imbalanced distribution of
classes(churn and nonchurn)with the high dimensional features,most
of the traditional rule-based methods may hard obtain acceptable pre-
diction accuracy.This paper proposes a new fuzzy-rule-based modelling
technique for churn prediction.Generally,the new prediction technique
heuristically discoveries a set of fuzzy rules,then uses this set of rules to
classify the unkown observation objects.Finally,the experiments were
carried out to evaluate the proposed modelling technique,based on5
datasets from UCI(University of California,Irvine)repository,and1
dataset of telecommunication.The experimental results show that the
proposed approach can achieve acceptable prediction accuracy and out-
performs the existing rule-based methods.
Keywords:Classi?cation,Fuzzy Rules,Hill-climbing Search,High Or-
der Rules
1Introduction
Service companies in telecommunications su?er from a loss of valuable customers to competitors;this is known as customer churn.In the last a few years,there have been many changes in the telecommunication industry,such as,new ser-vices,technologies and the liberalizations of the market increasing competition. Recently,data mining techniques have been emerged to tackle the challenging problem of customer churn in telecommunication service?elds[1,2,3,4].As one of the important measures to retain customers,churn prediction has been a concern in telecommunication industry and research.
As one of the important phases in churn prediction process,classi?cation can directly a?ect the prediction results.Many modelling techniques have been successfully used for the classi?cation,including Decision Tree(DT),Logistic
2Bingquan Huang,Ying Huang,M-T.Kechadi
Regression(LR),Arti?cial Neural Networks(ANN),Naive Bayes(NB)and Sup-port Vector Machine(SVM).However,they still have some limitations.Most of the modelling techniques provide either prediction probability or classi?cation rules.For example,NB,LR[5,6,7],ANN[8,9]and SVM[10,11]can only provide the probability of classi?cation,whereas they do not produce decision rules.DT [12,13,14]can produce rules,however,the probability of classi?cation cannot be obtained from DT model[15].
A number of Genetic Algorithm(GA)based classi?cation models also have been reported in[16,17,18].These models can be e?cient.However,they have high computational complexity.Recently,DMEL(Data Mining by Evolutionary Learning)[15],a GA based prediction approach was proposed.This technique not only provides classi?cation rules,but also gives the probability of classi?ca-tion.It was shown that DMEL is a very promising modelling technique.
This paper proposes a new classi?cation method which is based upon fuzzy-rule induction.It can not only classify samples by a set of generated rules, but also it is able to produce the prediction probability.Besides,this approach is capable of dealing with high-dimension and imbalance data set.There are three main procedures for implementing the new algorithm:Firstly,normalizing all values of each attribute to a speci?ed range,then fuzzy partitioning each attribute to several intervals.Secondly,generating rules by applying the heuristic of hill-climbing and pruning unimportant and redundant https://www.doczj.com/doc/3618712759.html,st but not least,predicting one of the categories to which a test case belongs according to the fuzzy rules.In the experiments,Lift Curve and AUC were applied to demonstrate the performance of algorithm.
The rest of this paper is organized as follows:Section2introduces the re-lated work.Section3describes the details of the proposed algorithm.Section 4provides the experimental results and discussion.The conclusion and future work are given in Section5.
2Related Work
Rule-based classi?cation is a well known domain.One of the main characteristics of rule induction is that they are more transparent and understandable than some other training models.A signi?cant amount of research e?ort has been put into this domain.In a rule-induction process,usually two strategies are applied:general-to-speci?c(top-down)and speci?c-to-general(bottom-up).Due to that the strategy of general-to-speci?c is more widely applied in rule learning, this section will introduce several approaches by using it,besides examples of rule induction by applying the strategy of speci?c-to-general can be found in [19,20,21].
Among all the di?erent methods for constructing the rules,decision tree based algorithms are the most popular[12,13,14].C4.5has proved to be one of the most well known among all tree-based classi?ers.A tree is constructed using the method of divide-and-conquer.C4.5starts to search an attribute with the highest information gain,and then the data is partitioned into classes according
Bingquan Huang,Ying Huang,M-T.Kechadi3 to this attribute.Each sub-tree represents a class,and will be recursively further partitioned using the same policy until some stopping criteria are satis?ed,such as the leaf node has been reached etc.The rules can be obtained by traversing each branch of the tree.The details of this algorithm can be found in[12].
CN2[22]works in an iterative way,each iteration searches one rule that covers(see section3.2for the de?nition)a large number of training cases by making use of beam search method.Entropy and likelihood ratio are used to evaluate the quality and the signi?cance of the rules.Rules are ranked in accor-dance with the quality of rules.When classifying a test case,CN2goes through each rule until?nding a rule that covers the case,then assigns the class label of that rule to the case.
[23]describes a genetic-algorithm-based method for selecting a small number of signi?cant fuzzy if-then rules to construct a compact classi?cation system. The performance of this method is examined by computer simulation on the irish data set.This method generate a set of fuzzy rules and optimize rules by applying genetic algorithm,however,it is di?cult to treat data sets with high-dimension attributes due to the expensive time complexity.
DMEL is a genetic classi?cation technique.The DMEL classi?er consists of a set of labelled rules which were found by rule induction technique.DMEL starts by generating first?order rules,then high?order rules can be constructed iteratively by randomly combining lower order rules.Given a sample without labels,the DMEL classi?er applies the rules to model the sample and makes the classi?cation decisions.DMEL has been shown to be an e?cient classi?er.The details of this technique can be found in[15].
3Algorithms
Although experimental results from[15]showed that DMEL is e?cient at detect-ing churners,there is still improvement that can be done for better classi?cation accuracy.The objective in designing the new classi?cation system is to improve the classi?cation accuracy.The proposed approach does so by considering several aspects:Firstly,fuzzy logic technique was applied to partition each attribute into several interval;Secondly,it applies di?erent heuristic methods for constructing classi?cation rules;Thirdly,constraints were introduced to get rid of bad and redundant rules.
3.1Fuzzy Partition
In this section,we brie?y talk about the partition for each attribute based on fuzzy logic.Fuzzy logic emerged as a consequence of the proposal of fuzzy set theory by Lot?Zadeh[24]in1965.In contrast with Crisp Logic,which has a binary sets,fuzzy logic variables may have a value that ranges in degree between 0and1.The degree,which represents how much extent dose an attribute value belongs to a subset of fuzzy partition,is managed by a membership function,such as triangle,trapezoid and exponential etc.In this paper,the classical triangle is
4Bingquan Huang,Ying Huang,M-T.Kechadi
applied as the membership function.Give a training set D={f1,...,f i,...,f n} where f i denote the i th attribute and n is the number of attributes in the whole feature space.Suppose each attribute in the feature space is partitioned into m fuzzy subsets S={s1,...,s i,...s m},s i represents the i th subset.The membership function is expressed as follows:
μi=max{1?|x
F
?(i?1)|,0}(1)
where
F=
1
m?1
(2)
Fig.1.An example of fuzzy partition for an attribute.
Figure1shows an example of fuzzy partition for an attribute,the vertical axis represents the degrees of membership,and the horizontal axis represents the values of the attribute,each instance has a di?erent attribute values that corresponds to a diverse degree of membership,the number of subsets was set to be5in this example.In order to generate rules based on the fuzzy partition of attributes,all attribute values in both the training and testing data set are supposed to be normalized to[0,1].
3.2Basic Concepts
The Rules are of the following form:IF antecedent THEN consequence,where antecedent describes a conjunction of a certain number of
Bingquan Huang,Ying Huang,M-T.Kechadi 5
let V i and e be the value of attribute f i and the total number of
The total degree of membership of the rules can be calculated by :
T DM =μ1·...·μi ...·μe ;(3)
where μi is the degree of membership which expresses the extent that V i belongs to s j .
Another basic concept of rule induction is Cover ,which expresses the case that a sample satis?es the antecedent part of the rule.A rule Correctly Covers a sample when it correctly predicts the covered sample.The task of rule induction is to build a model by generating a set of rules,which either do or do not Cover cases.
general rule and specific rule are two important https://www.doczj.com/doc/3618712759.html,ually,a rule with a smaller number of
3.3Rule Learning
The new prediction algorithm explores the strategy of general-to-speci?c,so it starts by generating the most general rules,?rst-order rules,then iteratively constructing higher-order rules based on the lower-order rules.Algorithm 1il-lustrates the process of learning rules.SET rules is the set of generated rules,k denotes the number of orders,and Max ?num ?order denotes the maximal number of orders.
Algorithm 1New Algorithm(training data)
1:
SET rules ←φ;2:
k ←1;3:
SET rules ←F irst ?order ?rules (training data);4:
while k <=Max ?num ?order do 5:
high ?order ←Higher ?order ?rules (lower ?order ?rules );6:
SET rules ←SET rules high ?order ;7:
k ++;8:
end while 9:return SET rules
6Bingquan Huang,Ying Huang,M-T.Kechadi
First-order Fuzzy Rules Algorithm2illustrates the process of generating the?rst-order fuzzy rules.In order to avoid information loss,all possible com-binations of
Algorithm2F irst?order?rules(training data)
1:F irst?order?rules←φ;
2:for i=1;i<=num?attribute;i++do
3:for j=1;j<=num?subset;j++do
4:for k=1;k<=num?class;k++do
5:if P runing(rule ijk)returns false then
6:F irst?order?rules←F irst?order?rules
rule ijk;
7:end if
8:end for
9:end for
10:end for
11:return F irst?order?rules
High-order Rules Higher-order rules are iteratively constructed based on the previous set of lower-order rules.The second-order is grounded on the first?order, the third?order is based on the second?order.Generally,(n?1)?order rules are the base for the n?order rules.Algorithm3illustrates the procedure of constructing a set of high?order rules.Positive lower order rules,which have the positive prediction,and negative lower order rules,which reversely have the negative prediction,can be separately obtained from the lower?order?rules.In algorithm3,item is used to de?ne a
Bingquan Huang,Ying Huang,M-T.Kechadi 7
applied.
W RA =Num ?(a )Num ?total ×(Num ?(a,c )Num ?(a )?Num ?(c )Num ?total )(4)
where Num ?(a )and Num ?(c )represent the number of data cases that are cov-ered by the antecedent and the consequence of a rule,respectively.Num ?(a,c )is the number of cases correctly covered by a rule,and Num ?total denotes the total number of cases in the training set.Line 9expresses that only rules with exclusive quality can be added.
Algorithm 3Higher ?order ?rules (a set of lower ?order ?rules )
1:
all ?rules ←φ;2:
positive ?items ←get all exclusive items from positive ?lower ?order ?rules ;3:
negative ?items ←get all exclusive items from negative ?lower ?order ?rules ;4:
accuracy ?list ←φ;5:
for i =1;i <=num ?positive ?rules ;i ++do 6:
high ?rule ←hill ?climbing (positive i ,positive ?item );7:
if P runing (high ?rule )returns false then 8:
W RA i ←calculate the W eighted ?Relative ?Accuracy for high ?rule ;9:
if W RA i is not in the list of accuracy ?list then 10:
all ?rules ←all ?rules high ?rule ;11:
accuracy ?list ←accuracy ?list W RA i ;12:
end if 13:
end if 14:
end for 15:
do the same work to generate negative ?rules ;16:
all ?rules ←all ?rules negative ?rules ;
17:return all ?rules ;
Algorithm 4illustrates the process of generating one high ?order rule.Let accuracy ?list ,which is a list consisting of di?erent WRA values,be empty in the beginning.Basically,one piece of high ?order is based on one piece of lower ?order rule and a new item ,which has been mentioned earlier.If the new item does not occur in the antecedent of the lower ?order rule,then one ?high ?order can be built by combining one ?lower ?rule and the new item .Meanwhile,the WRA value of the newly built one ?high ?order is stored in the accuracy ?list .This algorithm returns one of the high ?order rules with the highest WAR value.
3.4Pruning Rules
The number of generated rules can be huge,to make the classi?cation e?ective,we must prune bad rules with insigni?cant or noisy information.Several criteria are used for pruning https://www.doczj.com/doc/3618712759.html,ually,χ2statistic test is used to test if there exists
8Bingquan Huang,Ying Huang,M-T.Kechadi
Algorithm 4hill ?climbing (one ?lower ?rule ,all ?low ?items )
1:
accuracy ?list ←φ;2:
for i =1;i <=num ?low ?items ;i ++do 3:
if one ?lower ?rule does not include item i then 4:
one ?high ?rule ←combination of one ?lower ?rule and item i ;5:
W RA i ←calculate W eighted ?Relative ?Accuracy for one ?high ?rule ;6:
accuracy ?list ←accuracy ?list W RA i ;7:
end if 8:
end for 9:
BEST ←one of high order rules having the highest WRA;10:return BEST
strong relative relationship between two attributes.In this paper,χ2,which can be calculated by Eq.(5),is applied to decide whether a rule is signi?cant which means all the attributes occurring in antecedent part of the rule correlate to the consequence.
χ2
=(O ?E )2
E
(5)where O and E are the observed and expected frequency,which can be expressed by Eq.(6):O =Num ?(a,c )E =Num ?(a )×Num ?(c )
Num ?total (6)
In addition to χ2,Support and Confidence are another two essential factors when deciding if a rule should be pruned or not.Support and Confidence can be calculated by Eq.(7).
Support =Num ?(a,c )
Num ?total Confidence =Num ?(a,c )
Num ?(a )(7)
Algorithm 5illustrates the pruning method.αis the critical value for χ2signi?cance test,and it is set to be 3.84in this research.minS and minC are two threshold values,which de?ne the minimal value of support and con?dence,respectively.In this research,the minimum value of support and con?dence were set to be 0.01and 0.5,respectively.A rule can be retained if it satis?es all the threshold values.
3.5Classi?cation
This section discusses how to classify cases based on a number of classi?cation rules.Rules consist of two categories,churn and non ?churn .Two prediction models can be built based on all churn rules and all non ?churn rules,respec-tively.In order to estimate the importance of each rule,we rank rules in each model based on a set of principles,which are also used in [26]and can be de-scribed as follows:
Bingquan Huang,Ying Huang,M-T.Kechadi9
Algorithm5Pruning(rule)
1:F lag←true;
2:if chi-square statistics>αAND Support?rule>minS AND Confidence?rule> minC then
3:F lag←false;
4:end if
5:return Flag;
1.If confidence?1>confidence?2,then rule?1has higher priority than
rule?2;
2.If confidence?1=confidence?2and support?1>support?2,then rule?1
also has higher priority than rule?2;
3.If confidence?1=confidence?2,support?1=support?2and rule?1is
more general than rule?2,then rule?1has higher priority than rule?2.
After ranking,each rule has a position in each prediction model,one is more important than another rule if it has a better position than the other.We de?ne the signi?cance of each rule as follows:
Sig?level=Num?rules?position
Num?rules
×T DM(8)
where Num?rules is the number of rules,position denotes the rule index of the ranked rule set,refer to Equation3for calculating T DM.
Therefore,in a prediction model,the Sig?level of the most important rule is
1×T DM,while that of the least important one is T DM
Num?rules .To classify a case,
?nding all covered rules from the churn model and non?churn model,respec-tively,if the sum Sig?level of covered rules in churn model is greater than that of non?churn model,then assign churn to the case as the class label,otherwise, non?churn will be assigned.Apart from the mentioned two cases,if the current data case had the same sum Sig?level for churn model and non?churn,the majority class label should be assigned.
Moreover,the imbalance data set is a big challenge for data mining tech-niques.Therefore,for data set with the characteristic of imbalance,a threshold, which is used to adjust the weight of majority and the minority class,is better to be applied.In order to obtain a proper threshold value,we set a list of real numbers that are in the range of[0.9,1],the numbers are iterative added by 0.01up to the maximum value1.The5-fold cross-validation is applied on the training set,recording an AUC value each time by using the proposed fuzzy-rule prediction approach,?nally the appropriate threshold comes to the one which generated the maximum AUC.Therefore,when predicting an given instance, Equation8need to be replaced by the following one:
Sig?level=Num?rules?position
Num?rules
×T DM×threshold(9)
10Bingquan Huang,Ying Huang,M-T.Kechadi
In order to have the idea of which customers are more likely to churn,the new method scores each customer according to the sum of its Sig?level.If one is predicted to churn,then this customer is scored the sum of Sig?level of the churn model.Otherwise,the sum of Sig?level in non?churn model will be used to score that individual.Customers are sorted separately according to the prediction,cases predicted to churn are sorted in descending order,while cases predicted to stay(non?churn)are sorted in ascending order.For both of the sorted lists,customers close to the top are more likely to become churners.
4Experiments and Discussion
Experiments were conducted based on?ve UCI repository datasets[27]and a telecommunication dataset of the Ireland Telecoms[28].The datasets are de-scribed in Table1.Each of these datasets was equally divided into one training dataset and one testing dataset,they have the same distribution.
Table1.Dataset descriptions
Dataset Names#Instances#attributes Class Distribution
German Credit Data(UCI)20002030%:70%
Heart Disease(UCI)5401344.44%:55.56%
Brest Cancer1(UCI)5703137.19%:62.81%
Brest Cancer2(UCI)2003424%:76%
Japan Credit Screening(UCI)5691444.5%:55.5%
Telecommunication3256122 6.08%:93.92%
4.1Evaluation
In order to evaluate the classi?cation performance,the Lift Curve[29]and the area under Lift Curve(AUC)[30]are used.The Lift Curve measure is calculated by scoring each data case,sorting cases according to their scores,calculating two parameters that indicate the percentage of samples and the percentage of True Positive(true churn in this paper),respectively.The two parameters can be calculated by Eq.(10):
X=Num?Samples
Num?T otal
Y=
Num?T rue?P ositive
Num?T otal?P ositive
(10)
where Num?Samples,Num?T otal,represent the number of sample data, the total number of testing cases,respectively.Let Num?T rue?P ositive and Num?T otal?P ositive be the number of samples that have been identi?ed as positive,and the total number of positive cases in the whole testing dataset, respectively.
Bingquan Huang,Ying Huang,M-T.Kechadi 11
Occasionally,it is di?cult to use lift curve to evaluate the detected percent-ages from di?erent prediction modelling techniques or di?erent feature subsets of data.To overcome this problem,[30]calculates the area under curves by the following equation:
AUC =S 0?n 0×(n 0+1)×0.5
n 0n 1(11)
where S 0is the sum of the ranks of the class 0(churn)test patterns,n 0is the number of patterns in the test set which belong to class 0(churn),and n 1is the number which belongs to class 1(nonchurn).The details of AUC can be found in [29,30].
(a)(b)
(c)(d)
https://www.doczj.com/doc/3618712759.html,parison of Lift Curves between traditional and new modelling techniques,using di?erent datasets:(a):telecom,(b):heart,(c):credit,(d):japan
4.2Results and Discussion
In order to reduce costs,telecommunication operators randomly select a small amount of customers to o?er special consideration or attractive services.The objective is to ?nd a relative small amount of customers that are the relative more potential churners,by using lift curve technique.Figure 2shows the lift curves as the experimental results by using the di?erent modeling techniques (including the new proposed technique,DMEL,logict regression,MLP and Decision Tree
12Bingquan Huang,Ying Huang,M-T.Kechadi
(a)(b)
https://www.doczj.com/doc/3618712759.html,pare AUC values among di?erent models,using di?erent datasets (C4.5)),based on 4di?erent datasets.Sub-?gure 4.1represents the lift curves when a telecom dataset was used.The rest of sub-?gures are the ones when we used the three 3arti?cial datasets of the uci repository [27].
In each ?gure of Figure 2,the horizontal axis represents the percentage of selected customers,while the vertical axis is the ratio (percentage)between the number of the churners and the number of customers.In the ?gures,the di?erent colours represent the di?erent modelling techniques used in the experiments.The red curve in each sub-?gure is closest the top-left corner and the area under it is the largest.It means that the new proposed technique more accurately predicted the churners than other techniques.We can clearly conclude that the results based on the proposed approach are more e?ective than both the existing modelling techniques (DMEL,MLP,regression and tree-based algorithm).
There are a number of curves in each sub-?gure if Figure 2.We might have di?cult to identify which curve is better.Therefore,we used the AUC tech-nique to calculate the area under the curve for the evaluation of the modelling techniques.Figure 3shows the AUC values,which are produced upon diverse datasets and di?erent modelling algorithms.In each ?gure of Figure 3,the x-axis represents di?erent modelling tchniques and the y-axis represents the the values of AUC.In addition,the di?erent curves with di?erent colours represent di?erent datasets used in the experiments.Figure 3(a)shows the AUC values when all customers were selected.Figure 3(b)shows the AUC values when the half of customers were selected.From Figure 3(a)or Figure 3(b),the new pro-posed technique obtained the largest AUC value,when the same one dataset was used.Thus,the result of AUC measurement is consistent with the one of Lift Curve.That is,the AUC ?gures also represent the new proposed fuzzy-rules modelling techniques is more e?ective than the existing algorithms for customer churn prediction.
5Conclusion and future works
In this paper,we proposed a new classi?cation algorithm which is based on fuzzy rule learning.The experiments were conducted on a telecommunication
Bingquan Huang,Ying Huang,M-T.Kechadi13
data sets and5arti?cial data sets.Lift Curve and AUC techniques are applies to evaluate the performance of the algorithms.The experimental results show that the proposed classi?cation system is more e?cient than DMEL and tree-based methods.
The classi?cation results slightly vary when setting di?erent values for the order to generate rules.Thus,it is necessary to?nd a proper threshold to control the number of high order rules rather than setting the size arbitrarily.In addition, the minimum values of support and con?dence should also be considered more in the future since these will impact on the e?ectiveness of rule pruning. References
1. C.Wei and I.Chiu.Turning telecommunications call details to churn prediction:
a data mining approach.volume23,pages103–112,2002.
2.Shin-Yuan Hung,David C.Yen,and Hsiu-Yu Wang.Applying data mining to tele-
com churn management.Expert Systems with Applications,31:515–524,October 2006.
3. B.Q.Huang,M-T.Kechad,,and B.Buckly.Customer churn prediction for broad-
band internet services.In M.Bohanec,B.Kasek,https://www.doczj.com/doc/3618712759.html,vrac,and D.Mladenic, editors,Data Warehousing and Knowledge Discovery,Lecture Notes in Computer Science,volume5691,pages229–243.Springer Berlin Heidelberg,August2009.
4. B.Q.Huang,M-T.Kechadi,and B.Buckley.A new feature set with new window
techniques for customer churn prediction in land-line telecommunications.Expert Systems with Applications,37:3657–3665,2010.
5.David Williams,Xuejun Liao,Ya Xue,and Lawrence Carin.Incomplete-data
classi?cation using logistic regression.In In ICML,pages972–979.ACM Press, 2005.
6.Paul Komarek and Advised Andrew Moore.Logistic regression for data mining
and high-dimensional classi?cation.Technical report,2004.
7.Lemeshow S Hosmer D.W.Applied logistic regression.
8.William D.Penny,Stephen J.Roberts,and London Sw Bt.Bayesian neural net-
works for classi?cation:How useful is the evidence framework?,1998.
9. A.Bernatzki,W.Eppler,and H.Gemmeke.Interpretation of neural networks for
classi?cation tasks.In Proceedings of EUFIT,pages1420–1424,1996.
10.Dinggang S Yiqiang Z.Design e?cient support vector machine for fast classi?ca-
tion.IEEE Transactions on Neural Neworks,11:124–136,January2000.
11. C.J.C.Burges.A tutorial on support vector machines for pattern recognition.Data
mining and knowledge discovery,2(2):121–167,1998.
12.J.Ross Quinlan.C4.5:Programs for Machine Learning(Morgan Kaufmann Series
in Machine Learning).Morgan Kaufmann,1edition,January1993.
13.J.R.Quinlan.Improved use of continuous attributes in c4.5.Journal of Arti?cial
Intelligence Research,4:77–90,1996.
14.Wenliang Du and Zhijun Zhan.Building decision tree classi?er on private data,
2002.
15.W.Au,C.C.Chan,and X.Yao.A novel evolutionary data mining algorithm with
applications to churn prediction.IEEE Transactions on Evolutionary Computa-tion,7:532545,2003.
14Bingquan Huang,Ying Huang,M-T.Kechadi
16.James D.Kelly.A hybrid genetic algorithm for classi?cation.In In Proceedings of
the Twelfth International Joint Conference on Arti?cial Intelligence(pp.645–650, pages645–650.Morgan Kaufmann,1991.
17.Wojciech Kwedlo and Marek Kretowski.Discovery of decision rules from databases:
An evolutionary approach.In In Principles of Data Mining and Knowledge Discov-ery.2nd European Symposium PKDD’98.Springer LNCS1510,pages370–378.
Springer-Verlag,1998.
18.Min Pei,Erik D.Goodman,William F.Punch Iii,and Ying Ding.Genetic algo-
rithms for classi?cation and feature extraction.In Annual Meeting,Classi?cation Society of North America,1995.
19.Pedro Domingos.Rule induction and instance-based learning:A uni?ed approach.
pages1226–1232.Morgan Kaufmann,1995.
20.Pedro https://www.doczj.com/doc/3618712759.html,ing partitioning to speed up speci?c-to-general rule induc-
tion.In In Proceedings of the AAAI-96Workshop on Integrating Multiple Learned Models,pages29–34.AAAI Press,1996.
21.Pedro Domingos.E?cient speci?c-to-general rule induction.AAAI Press,1996.
22.Peter Clark and Tim Niblett.The cn2induction algorithm.In Machine Learning,
pages261–283,1989.
23.Naohisa Yamamoto Hisao Ishibuchi,Ken Nozaki and Hideo Tanaka.Selecting
fuzzy if-then rules for classi?cation problems using genetic algorithms.volume3, pages260–270,1995.
24.L.A.Zadeh.Fuzzy https://www.doczj.com/doc/3618712759.html,rmation and Control,8:338–353,1965.
25.Peter Flach Ljupo Todorovski and Nada Lavra.Predictive performance of weighted
relative accuracy.In Principles of Data Mining and Knowlege Discovery,Lecture Notes in Computer Science,volume1910/2000,pages182–211.Springer Berlin Heidelberg,2000.
26.Class association Rules,Wenmin Li,Jiawei Han,and Jian Pei.Cmar:Accurate
and e?cient classi?cation based on multiple class-association rules,2001.
27.https://www.doczj.com/doc/3618712759.html,/ml/datasets.html.
28.http://www.eircom.ie/cgi-bin/bvsm/mainPage.jsp.
29.Miha Vuk.Roc curve,lift chart and calibration plot,2006.
30.Andrew P.Bradley.The use of the area under the roc curve in the evaluation of
machine learning algorithms.Pattern Recognition,30:1145–1159,1997.
掌握Vegas快捷键的好方法 掌握Vegas快捷键的好方法DIY Sony Vegas键盘 转lcfxingfu 下面是我常用快捷键 J 反向搜索 K 暂停 L 正向搜索 Enter 播放/暂停(光标就地停止) Space 播放/停止(光标返回播放点) S 截断 M 打点 G 组合 U 解组 I 标记头 O 标记尾 R 标记视频区域 N 标记音频区域 ` 折叠轨道开/关 Tab 在标题栏和时间线之间跳转 Ctrl+T 裁切[X]保留循环区域 Ctrl+X 裁切]X[剪掉循环区域Shift+Ctrl+W 开/关图示提高素材读取速度 Shift+Ctrl+I 开/关名示素材标记文件名 Shift+Ctrl+C 开/关图标素材段落密集时可防止错点击，尤其是在单纯的裁剪阶段Shift+Ctrl+T 开/关转场时间显示 有一些是不用标注的 1.标准的windows键 Ctrl+Z返回 Ctrl+A全选 Ctrl+S另存 Ctrl+N新建 Ctrl+O打开 Ctrl+C拷贝 Ctrl+V粘贴 2.非常形象的键 [、] 素材头、素材尾 <、> 区域头 →、←方向键移动光标（ALT+→、Alt+←光标移动单桢同F3、F9） ↑、↓方向键缩放轨道 →、← (4、6)键移动素材 ↑、↓ (8、2)键跳动素材
Insert 插入关键桢 Delete 删除 Home 光标到起始点 End 光标到末尾点 Up 光标整数左移 Down 光标整数右移 3.特殊用途的快捷键 如：Alt+D 再按2 恢复默认窗口 Vegas快捷键大集合 Vegas Video 快捷键1 Project file 方案文件 命令热键Create new project 建立新方案Ctrl+N Exit Vegas 退出Vegas Alt+F4 Open existing project 打开已存在的方案Ctrl+O Project properties 方案特性Alt+Enter Save project 保存方案Ctrl+S Window view 调用视窗视图 命令热键 Set focus to track view Alt+0 Explorer 导出Alt+1 Trimmer 修剪器Alt+2 Mixer 混音台Alt+3 Edit Details 编辑细节Alt+4 Media Pool 媒体存储表Alt+5 Video Preview 视频预览Alt+6 Transitions 转变Alt+7 Video FX 视频特技Alt+8 Text/Backgrounds 文本/背景（字幕）Alt+9 Plug-Ins 插件Ctrl+Alt+1 Next window 下一个视窗F6或Ctrl+Tab Previous window 前一个窗口 Shift+F6 或
一．Delay延迟效果器。 延迟效果器能够在声音播放间隔一定时间后，再次播放同样的声音，以产生拖延和重复感。 它一般有以下几个参数： 1．干声输出（dry out）:是指原来的声音输出的音量大小。 2．延迟输出（delay out）:是指声音再次回放时的音量大小。 3．延迟时间（delay time）:是指原来的声音和回放时的声音之间的时间间隔。 4．衰减时间（decay time）:是指声音从原始声音后的第一次回声开始，回声的音量依次减弱，直到最后一次回声的完全消失所用的时间。时间越长重复回声的次数就越多，反之，越少。 二．Chorus 合唱效果器。 合唱效果器是指在原来的声音基础上叠加一些由计算机产生的类似的声音样本，并和原乐器叠加，使单个乐器听起来象多个乐器一样，更有层次感。 合唱效果器一般有以下几个参数： 1． Input gain(输入增益)：用来控制输入音量的大小。输入增益越大，合唱的效果越明显；输入增益越小，合唱的效果越不明显。 2． Dry out (干声输出)：用来调节原来的声音输出的音量大小。最大调节可以保持原来声音的面貌。 3． Chorus out(合唱输出)：用来调节产生的合唱效果的音量大小。也就是调节在原来声音基础上产生的新的声音样本的音量大小。 4． Chorus size(合唱空间)：用来控制合唱空间的大小。空间越大，共鸣越强，合唱效果就越明显。 5． Feedback(声音反馈)：是用来调节产生声音的反馈长短。如果加大这个值，可以明显地听出原有干声基础上的合唱声音变长了。如果再加大这个值，就能听到明显的金属共鸣声。 6． Chorus out delay(合唱延迟时间)：是指基础声音播放后，再次播放产生的新声音所间隔的时间。通常在0.1~100ms之间，超过50ms后，延迟听起来就非常明显了。切记，不可把这个延迟时间做为延迟效果器使用，应严格控制。
QuartusII软件的使用方法 冯海芹编 四川托普信息技术职业学院 电子与通信系 QuartusII的设计流程 QuartusII软件的使用方法 一、设计输入 1.建立工程 任何一项设计都是一项工程（Project），都必须首先为此工程建立一个放置与此工程相关的所有设计文件的文件夹。此文件夹将被EDA 软件默认为工作库（Work Library）。一般，不同的设计项目最好放在不同的文件夹中，而同一工程的所有文件都必须放在同一文件夹中。 首先建立工作库目录，以便存储工程项目设计文件。在D盘下新建文件夹并取名Mydesign。双击QuartusII软件启动图标，即可启动QuartusII软件，启动界面如图1-2所示。
使用New Project Wizard 可以为工程指定工作目录、分配工程名称以及指定最高层设计实体的名称，还可以指定要在工程中使用的设计文件、其他源文件、用户库和EDA 工具，以及目标器件系列和具体器件等。在此要利用“New Preject Wizard”工具选项创建此设计工程，并设定此工程的一些相关的信息，如工程名、目标器件、综合器、仿真器等。 (1)打开建立新工程管理窗。选择菜单File→New Preject Wizard 命令，即弹出“工程设置”对话框（图1-3），以此来建立新的工程。 (2)在单击图1-3后，出现了设置工程的基本信息，如图1-4所示。单击此对话框最上一栏右侧的“… ”按钮，可以选择工程存放在硬盘上的位置，此例中将工程放在D盘Mydesign文件夹下。这三行的第一行的d:\Mydesign表示工
程所在的工作库文件夹；第二行的half_add 表示此项工程的工程名，工程名可以取任何其他的名，也可直接用顶层文件的实体名作为工程名，在此就是按这种方式取的名；第三行是当前工程顶层文件的实体名，这里即为half_add。 （2）将设计文件加入工程中。单击图1-4中的Next 按钮，弹出对话框如图1-5所示，在对话框中单击File name 右侧的“… ”按钮，可以将与工程相关的所有VHDL 文件（如果有的话）加入进此工程，此工程文件加入的方法有两种：第1 种方法是单击“Add … ”按钮，从工程目录中选出相关的VHDL 文件；第2 种方法是单击Add All 按钮，将设定的工程目录中的所有VHDL 文件加入到工程文件栏中。如果还没有建立VHDL文件，就直接点击“Next”即可。
欧美流行音乐之管见 ----R&B 内容摘要：R&B的简介，起源，最初R&B的风格，R&B的大众化，R&B音质多变化，现 代节奏的布鲁斯 关键词： 正文：简介“节奏布鲁斯”（R&B）在20世纪40年代中期出现的时候，它甚至还没有名字。但这个词才一出现，它就迅速广泛地传播开去。时至今日，R&B已经成了黑人流行音乐的代名词。 R&B的全名是Rhythm & Blues，一般译作“节奏怨曲"。广义上，R&B可视为“黑人的流行音乐”，它源於黑人的Blues音乐，是现今西行流行乐和摇滚乐的基础，，曾界定R&B为所有黑人音乐——除了Jazz和Blues之外，都可列作R&B——可见R&B的范围是多么的广泛。近年黑人音乐圈大为盛行的Hip-Hop和RAP都源於R&B，并且同时保存着不少R&B 成分。 早期的摇滚乐就是以R&B为基础的，它是由受流行音乐影响的“乡村和西部音乐”延展而来。R&B不仅仅是在布鲁斯和摇滚乐之间的一种重要的过渡音乐，它还是布鲁斯、灵魂乐、摇滚、爵士乐之间最重要的音乐分支。节奏布鲁斯是爵士乐和蓝调的综合体（a combination of jazz and blues）。 当然，布鲁斯无疑是R&B的一个重要组成部分，但爵士乐元素也同等重要。 起源 起源——蓝调 R&B的前身，出现在40年代的美国黑人区。那时，负担得起多姿多彩娱乐的人不多，热爱音乐的黑人就喜欢茶余饭后聚集街头，以吉他和口琴等简单乐器玩奏音乐，一抒生活、工作、离乡别井等种种感受。这就发展出蓝调（Blues）——R&B的前身。随著录音带的发明，和越来越多的小型电台出现，蓝调音乐开始跳出黑人圈子，在芝加哥流行起来。不知是否因蓝调音乐说愁说得太久，乐迷都渴望听到节奏感强、调子节拍明快，内容又不太悲伤的蓝调歌曲。 最初开始流行的R&B风格通常是指“跳跃布鲁斯”（Jump Blues）。它不仅吸收了爵士乐里的摇摆节奏和以号为主的编配方式，而且吸收了布鲁斯里普遍使用的回复段与和声结构。在“跳跃布鲁斯”中，歌手的声音更加刺耳，节奏更快，乐器的演奏也不一样——钢琴的弹奏力度更大，萨克司的声音更加长而尖锐。 最重要和最具流行性的“跳跃布鲁斯”明星是路易斯·乔丹（Louis Jordan），他的唱片在黑人听众和白人听众里取得了同样的成功。 “跳跃布鲁斯”后来转变为几种不同的风格。那些被称作“大嗓门们”（Shouters）的演唱者在“大乐队”拘谨的演唱方式里加入了更具活力的“福音音乐”（Gospel music）和“灵魂乐”（Soul）。 50年代初期到中期，摇滚乐开始崭露头角。一些和R&B有着明显区别的音乐类型大显身手，它们依靠自身的力量对流行音乐文化发挥着不同的巨大影响。这些音乐在摇滚史上延续的影响力甚至超过了更早、更有爵士味的R&B。 音质多变化 不论歌声或乐器的音质（timbre），都会有很大变化。
Vegas快捷键大集合 Vegas Video 快捷键1 Project file 方案文件 命令 热键 Create new project 建立新方案Ctrl+N Exit Vegas 退出Vegas Alt+F4 Open existing project 打开已存在的方案Ctrl+O Project properties 方案特性Alt+Enter Save project 保存方案Ctrl+S Window view 调用视窗视图 命令热键 Set focus to track view Alt+0 Explorer 导出Alt+1 Trimmer 修剪器Alt+2 Mixer 混音台Alt+3 Edit Details 编辑细节Alt+4 Media Pool 媒体存储表Alt+5 Video Preview 视频预览Alt+6 Transitions 转变Alt+7 Video FX 视频特技Alt+8 Text/Backgrounds 文本/背景（字幕）Alt+9 Plug-Ins 插件Ctrl+Alt+1 Next window 下一个视窗F6或Ctrl+Tab Previous window 前一个窗口 Shift+F6 或 Ctrl+Shift+Tab Toggle focus between track list and timeline 在轨道表与时间线之间交替 Tab 在前一轨与下一轨之间切换 Alt+Shift+上/下箭头键 Vegas Video 快捷键2 Edit 编辑命令 命令 热键 Cut selection 切掉选择区Ctrl+X或Shift+Delete Copy selection 拷贝选择区Ctrl+C或Ctrl+Insert Paste 粘贴Ctrl+V 或Shift+Insert Paste Insert 插入粘贴Ctrl+Shift+V Paste repeat 重复粘贴Ctrl+B Delete selection 删去选择区Delete Mix to new track 混合到新轨Ctrl+M
蓝色的梦幻──Nomad Factory Blue Tubes Bundle 效果器组合的使用 蓝色，是一种充满了梦幻色彩的颜色。蓝色，总是能让我们想到深邃的蓝天、无边无际的大海。从太空望下来，我们的地球就是蓝色的。有人说，蓝色代表着包容和博爱，有人说，蓝色象征着宁静和自由，蓝色，仿佛是透过一丝阳光的海洋，蓝色，仿佛是……（读者叫道：大觉者！我们的牙快酸掉了！！！）哦好了好了，文章开篇语已凑够字数，接下来的这篇文章中，让我们一起来看一看，蓝色的声音是什么样子的！ Blue Tubes Bundle效果器，可以直译为“蓝色电子管组合”。顾名思义它是一套模拟电子管的效果器插件。目前，模拟电子管的插件有很多，比如Antares Tube、amplitube等等。电子管效果器主要是模拟电子管设备那种“暖”声，电子管所发出的光线往往也带给我们一种温暖的感觉。而我们今天讲的这套效果器，其电子管界面却是别出心裁地设计为蓝光，而它的声音，也确实很别致，功能之强大也绝非其他单独的电子管模拟插件所能相比，其资源占用却非常小，的确是一套非常好的效果器插件。 Blue Tubes Bundle效果器当前最高版本为2.0。一共包括15个效果器。这套VST效果器插件的安装非常简单，直接装到宿主下的Vstplugins文件夹里就OK了。在此就不再赘述了。在装上之后，我们就可以在宿主里找到它，以Nuendo3为例（如图）。
在这里我们可以看到这15个效果器。接下来我们来详细介绍一下这些效果器各自的特点。 首先说说第一个效果器，也就是BrickWall Limiter。这个效果器其实可以直译为“砖墙”或者“限制墙”，也就是说，它可以限制波形在出现峰值的时候不会产生爆音。实际上它已经使声音的波形变平滑了。在录制女高音、唢呐等等声压较大的声音时，这个效果器可以很好地派上用场。如下图，在波形的峰值上，BrickWall很好地控制住了声音。
PADS 9.3 Logic MainFrame Shortcuts Command Shortcuts Description Add Connection
欧美流行英文歌曲 A 1． My Heart Will Go On 我心永恒（铁塔尼克号） 2． Speak Solfty Love 轻声细语（教父） 3． Because I Love You 因为我爱你（走出非洲） 4． Beauty And The Beast 美女与野兽（美女与野兽） 5． Take My Breath Away 带走我的呼吸（比翼雄鹰） 6． All Out Of Love 失落的爱（谈情说爱） 7． Only You 只有你（罗马假日） 8． The One You Love 唯一的爱（时光倒退70年） 9． A Whole New World 崭新世界（阿拉丁） 10．The Power Of Love 爱的力量（回到未来） 11．Changing Partners 交换舞伴（钢琴课） 12．Nothing’s Gonna Change My Love For You 此情永不移（廊桥遗梦） B 1． Say You, Say Me 说你，说我（白昼）（百夜逃亡） 2． Love Story 爱情故事（爱情故事） 3． Hotel California 加州旅馆（三步杀人曲） 4． Love Is All Around 堕入爱河（四个婚礼和一个葬礼） 5． Don’t Cry For Me Argentina 阿根廷，别为我哭泣（贝隆夫人）6． When I Fall In Love 当我堕入爱河（西雅图夜未眠） 7． The Way We Were 往日情怀（俏郎君） 8． You Need Me 你需要我（偷窥） 9． Let It Be 顺其自然吧（刺激1995） 10．Lemon Tree 柠檬树（民歌） 11．Can You Feel The Love Tonight 今夜你感到爱了吗？（狮子王）12．LA COPA DE VIDA（The Cup Of Life）圣杯之光（’98世界杯主题曲）
影视后期制作 第一周 完善学生自己电脑的系统，并安装一系列相关专业软件。例如，Vegas，Pr，AE等等。 第二周 主要内容：影视相关的常识及Pr、Vegas的简单入门 周一 1.影视艺术的特征：画面与声音的艺术，利用视听觉营造真实感与现场感 ①声音在电影中的作用，声音的空间 ②画面：细节的真实 ③景观电影：视听觉的享受 ④未来的沉浸式电影：超越视听 2.后期与前期的关系 a.前期的某些不足可以在后期弥补，但是后期难度被加大了 b.后期充分的参与到前期制作中，有利于最终效果的完美 c.依赖于导演团队的整体把握能力 3.影视后期创作的流程： a.创建项目和导入素材 b.创建合成和排列图层 c.添加动画元素和特效 d.预览效果 e.渲染输出 4.AE的应用：电影、电视、多媒体、网络视频、手机视频（或其他手机设备视频）
和DVD编创行业中。 5.AE概览 ①动态的photoshop：遮罩、层、路径、滤镜等 ②和Adobe家族软件的结合良好 ③吸收了三维软件的关键帧、粒子、Z通道、灯光、等 ④支持节点的操作模式 ⑤支持maya的场景文件 ⑥支持rpf格式，诸如景深、空气透视、雾效就不一定在三维软件中制作了。 ⑦庞大的插件群：千余种 术语： 1.位图图形：像素数量是固定的 2.矢量图形：通过数学方程式产生 3.像素：是构成图形的基本元素 （越高位的像素拥有的色板越丰富，越能表达颜色的真实感） 4.分辨率：指屏幕图像的精密度，即显示器所能显示的像素的多少 5.色彩深度：是每个像素可显示出的色彩数 (单位：位 bit) 8位色彩→ 256色 16位色彩→中彩色 24位色彩→真色彩（百万色） （32位色彩与24位色彩在画面上没有什么区别，多出来的8位用来提现素材半透明的程度，也被称为Alpha通道） CMYK色彩模式：主要用于出版印刷领域 不用于视频编辑领域 AfterEffect CS５不支持此色彩模式的素材文
1、TC出品的效果器插件包。TC的这些VST效果插件一直都是被广泛使用. TC Compressor DeEsser压缩、消唇齿音效果器 Compressor压缩效果可以这样理解，就是把音频低的地方提升，把高的地方下压，以便让音频整体的音量更均匀，通过设置压缩的比例和起始时间以及释放时间，可以让一些比如低鼓、军鼓、BASS等乐器听起来感觉更有力，DeEsser我们一般翻译为消唇齿音效果，也有叫嘶声消除器的。它可以通过调整压缩、门限的参数来消除人声或乐器4KHZ到8KHZ之间的嘶声。比如唱歌时由口齿发出的唇齿音、箱琴在弹奏时发出的一些杂音。 中英文名词对照：Compressors/压缩、attack time/起始时间、threshold/门限、release time/释放时间、ratio/压缩 比例、SoftKNEE/拐点柔软度、Hard Time/硬的时间，SoftKNEE和Hard Time一起来设置拐点柔软度是硬还是软。De- Esser/嘶声消除。（其他品牌的压缩效果器都大同小异，不再重复解释，实际运用后面章节叙述）
TC Filtrator滤波效果器 简单来说Filtrator就是通过过滤某些频段和调整失真饱和度以及低频震荡来创造出一个全新的声音。TC Filtrator滤波 效果器主要分为：FILTER滤波模块、LFO低频震荡模块、DRIVE失真度模块并在ENV FOLLOWER模块中调整相关参数值。如 果你需要把你的声音变成一个外星人或者类似机器人的声音，想得到一种特殊的效果你不仿试试它，不过想用好可不是 那么简单。 中英文名词对照：Filter/滤波器、LFO/低频震荡器、Speed/速率、Division/分界点、Slope/倾斜值、Attack/起始时间 、HOLD/保持时间、DECAY/衰变时间（其他品牌的滤波效果器也是大同小异，不再重复叙述）
大键盘的加号和减号：调音 J 反向K 暂停L 正向 回车播放/光标就地暂停空格播放/光标返回播放点) S 截断M 打点G 组合U 解组I 标记头O 标记尾 R 标记视频区域N 标记音频区域`折叠轨道开/关 Tab 在标题栏和时间线之间跳转Ctrl+T保留循环区域 Ctrl+X剪掉循环区域 Shift+Ctrl+W 开/关提高素材读取速度 Shift+Ctrl+I 开/关标记素材名 Shift+Ctrl+C 开/关图标素材段落密集时可防止错点击 Shift+Ctrl+T 开/关转场时间显示 1.标准的windows键 Ctrl+Z返回Ctrl+A全选Ctrl+S另存Ctrl+N新建 Ctrl+O打开Ctrl+C拷贝Ctrl+V粘贴 2.非常形象的键 [、] 素材头、素材尾<、> 区域头→、←方向键移动光标（ALT+→、Alt+←光标移动单桢同F3、F9）↑、↓方向键缩放轨道→、←(4、6)键移动素材↑、↓(8、2)键跳动素材 Insert 插入关键桢Delete 删除Home 光标到起始 点End 光标到末尾点 Up 光标整数左移Down 光标整数右移
Alt+D 再按2 恢复默认窗口 （一）：逐帧找点： 1、按ALT健＋左右箭头健（左右搜索找点）。 2、按F9健正向搜索找点。 3、区域选择找点，在第一点按I健出现标记，然后选 择第二点按O健（清除找点按“转到开始”健，按I、O健）。 （二）、分割： 单击分割点，按S健分割成2段， （三）、剪切、删除（不含CTRL+X）： 1、按S健分割，选择被分割部分再按delete健，被分割部分删除。 2、选择区域按CTRL+T健，被选择区域以外的部分被剪切删除。 3、在修剪器窗口选择需要剪切的片段，然后在时间线编辑窗口点击插入点，按A健就可以剪切插入（可以移动定位）。 （四）、浏览、播放： 1、按空格健，再按停止。 2、按F12健，再按停止。 3、快速播放，左右拖动RATE杠杆（按J健左向、按L健右向，按L、J健不放开，加快速度，按空格健、K健停。 Vegas 命令热键 快捷键1 建立新方案Ctrl+N 退出Vegas Alt+F4 打开已存在的方案Ctrl+O 方案特性Alt+Enter 保存方案Ctrl+S
BOSS GT-6的中文说明书 第一章使用效果器演奏 导线连接方式 开启电源 调整输出音量 关闭电源 为连接装置设定输出方式(放大器;输出选择) 何谓音色 如何切换音色 使用gt-6选择音色 只切换阿拉伯数字(1-4) 切换数据库(bank 1-85)与阿拉伯数字(1-4) 显示屏的指示讯息 如果音色无法切换 第二章建立专属的常用音色 寻找与理想中类似的音色(ez tone) 使用旋钮调整音色 增加效果 设定效果音质(快速设定) 使用单一参数做更多细部设定 使用踏板效果(哇音器,踏板弯音) 哇音器(wah) 踏板弯音(pedal bend) 音色命名 改变效果连接顺序(effect chain) 第三章储存您自己制作的音色 写入(write) 复制音色 调换音色 第四章效果说明 前级放大器preamp/音箱speaker仿真器 过载/失真(overdrive/distortion) 延时(delay) 合声(chorus) 混响(reverb) 放大器(wah) wah(踏板哇音) aw(自动哇音) fw(修正哇音) eq(均衡器) fx-1(效果器-1) cs(压缩器) lm(限制器)
ac(木吉他仿真器) pic(拾音器仿真器) tr(颤音) sg(增音器) fb(回授器) afb(反回授器) frt(无音格仿真器) fx-2(效果器-2) ph(相位器) fl(飘忽器) hr(合音器) ps(移调器) pb(弯音踏板) 2ce(双和声器) pan(自动左右相位效果) vb(颤音效果) uv(独特相位器 sdd(短暂延时) hu(人声效果 rm(铃响仿真器 sl(断音器 ar(自动反复乐句效果 syn(吉他合成效果) seq(副均衡器) ns(杂音抑制器) master(主控参数) fv(脚控音量) 第五章表情与控制踏板的设定(踏板设定) 表情踏板的设定 表情踏板开关/控制踏板的设定 快速设定(quick setting) 编辑”快速设定” 利用gt-6踏板、外接踏板,与外接midi设备,来控制效果器使用外接表情踏板来控制哇音效果与弯音效果 第六章使用自订功能 进行”自订”前级放大器设定 进行”自订”过载/失真设定 进行”自订”踏板哇音设定 第七章 gt-6的便捷功能 快速音色音量调整(patch level旋钮) 使用踏板来开启与关闭效果器(manual mode) 切换到手动模式 选择用踏板切换开启与关闭的效果器 手动模式下,比较前级放大音色
一般来说我们用saber2007我们会用到两个应用，一个是saber simulator这个是标准用来仿真的，另一个是saber sketch这个是用来画电路图的，但是呢，现在我们就用一个应用saber sketch 它有个快捷键可以调用saber simulator 就是下图这个。PASS:不要懒网上教程一大堆 多搜搜吧。。。 点开前的界面↓。 点开后的界面↓。
上面就介绍到那，我们继续下一步，器件的查找和放置。
下面是如何查找器件↓。 下面这个是放置器件↓
器件的参数填写和一般的参数代表的意思，其一↓
（在这有必要的提醒下对于英语不好的同学还是安装 一个翻译软件比较好，我安装的是必应词典。） 器件的参数填写和一般的参数代表的意思，其二↓ 下面是saber不知道是那个版本的器件库介绍，我估计着是2006吧？但是我看了下都可以找到，我们就多了两个A开头的库（Aerospace和Automotive）。。。呃，貌似有点乱，你们也就将就着看了吧！！！ Saber Parts Gallery:Assembled by SubLater( 按中文意思归类，括号中是特殊部件和必要注释) ★Characterized Parts Libraries 特性元件 ├─DX ├─Diode 二极管(Zener 齐纳、Power 功率) ├─BJT 三极管(Darlington 达林顿、Power 功率、
Array 阵列) ├─JFET/MOSFET/ 功率MOSFET 场效应管 ├─SCR/IGBT ，Switch 模拟开关器件 ├─Analog Multiplexer 模拟多路开关 ├─OpAmp 运算放大器 ├─Comparator 比较器 ├─ADC 、DAC ├─Fuse 保险丝、Resettable Fuse 可复位保险丝(PPTC) ├─Inductor 电感线圈 ├─Transformer 变压器 ├─Motor 电机模型 ├─PWM 控制器、PFC 元件 ├─Schmitt Trigger 施密特触发器 ├─Sensor 传感器 ├─Timer 定时器 ├─Transient Suppressor 暂态抑制器 ├─Voltage Reference 电压参考给定 ├─Voltage Regulator 电压调节器 ├─SL ├─Diode 二极管(Zener 齐纳) ├─BJT 三极管(Darlington 达林顿)
vegas快捷键大全 vegas提供最优化HD剪辑的核心技术，同时也提升了回放的效能，加强对XDCAM工程的支持，全新的字幕技术，提升的剪辑功能，还有更多重大的更新和功能。Vegas Pro是面向所有专业人员的终极的多功能软件产品。他集合了Vegas Pro提供快捷的速度，强大的功能，和最大的创作效率。 ==========================Vegas Video 快捷键1=========================== Project file 方案文件命令热键 Create new project 建立新方案Ctrl+N Exit Vegas 退出Vegas Alt+F4 Open existing project 打开已存在的方案Ctrl+O Project properties 方案特性Alt+Enter Save project 保存方案Ctrl+S Window view 调用视窗视图命令热键 Set focus to track view Alt+0 Explorer 导出Alt+1 Trimmer 修剪器Alt+2 Mixer 混音台Alt+3 Edit Details 编辑细节Alt+4 Media Pool 媒体存储表Alt+5 Video Preview 视频预览Alt+6 Transitions 转变Alt+7 Video FX 视频特技Alt+8 Text/Backgrounds 文本/背景（字幕）Alt+9 Plug-Ins 插件Ctrl+Alt+1 Next window 下一个视窗F6或Ctrl+Tab Previous window 前一个窗口Shift+F6 或Ctrl+Shift+Tab Toggle focus between track list and timeline在轨道表与时间线之间交替Tab 在前一轨与下一轨之间切换Alt+Shift+上/下箭头键 ==========================Vegas Video 快捷键2=========================== Edit 编辑命令命令热键 Cut selection 切掉选择区Ctrl+X或Shift+Delete Copy selection 拷贝选择区Ctrl+C或Ctrl+Insert Paste 粘贴Ctrl+V 或Shift+Insert Paste Insert 插入粘贴Ctrl+Shift+V Paste repeat 重复粘贴Ctrl+B Delete selection 删去选择区Delete Mix to new track 混合到新轨Ctrl+M Open in audio editor 在音频编辑器中打开Ctrl+E Undo 恢复Ctrl+Z或Alt+Backspace Redo 重做Ctrl+Shift+Z 或Ctrl+Y Split 分裂（分割）S Trim/crop selected events 休整/修剪被选择事件Ctrl+T Rebuild Peaks 重新计算波形图F5 Normal tool 切换到普通工具Ctrl+D
前级效果器的调试方法 KTV包房中，一直以来回声混响的调校，主要依靠调音师所积累的经验和临场。极少有人在这个方面进行系统的原理分析，形成较为科学和系统的知道理论。QQ:19680308 Puresound?纯音实验室专门针对这一方面进行了深入研究，总结出回声混响使用的一些理论和技巧，在这里抛砖引玉，分享给各位合作伙伴。更希望能得到各路高手的指正和分享。首先我们来明确一下回声和混响的特征： 回声（Echo&Delay）：声波在传播过程中，碰到大的反射面（如建筑物的墙壁等）后反射回来，再度能够听到的声音，称之为回声。人耳能辨别出回声的条件是反射声具有足够大的声强，并且与原声的时差须大于0.1秒。回声的声音特征为：Echo跟Delay效果很类似，不同处在于Echo的音质会受到反弹物(如墙壁)材质、距离的影响；而Delay效果则纯粹将原音再拷贝，不影响音质。 混响（reverberation）：声波在室内传播时，要被墙壁、天花板、地板等障碍物反射和散射，每反射散射一次都要被障碍物吸收一些。这样，当声源停止发声后，声波在室内要经过多次反射和吸收，最后才消失，我们就感觉到声源停止发声后声音还继续一段时间。这种现象叫做混响。由于被多方向的反射面反射、散射以及部分吸收，混响的声音特征为带
有波浪感的虚化声音。 从上面对回声和混响的定义中，我们可以非常直观的得到回声和混响的声音特征：回声基本是原音的拷贝，而混响是虚化的声音。 在KTV演唱中，使用回声和混响的目的非常的明确：帮助演唱者进行演唱。而帮助演唱者获得较好的演唱感受，主要实现以下几个方面的辅助和补偿： 1、人声和音乐的融合； 2、营造较好的空间感； 3、对普通演唱者的人声增强，让演唱的声音相对饱满。 在目前的KTV行业中，目前大多数调音师最为常用的方式为重回声、轻混响。主要使用回声，稍许使用混响补充来达到以上的目的。 那么，之前我们对回声和混响的定义分析，采用这样的手段，是否能够达到我们满意的音效呢？下面我们详细的进行一些分析和讨论，来探讨实现以上三个方面，我们究竟该使用什么样的方法更加合理： 1、人声和音乐的融合 人声和音乐的融合是KTV演唱中的最基本要求，一旦人声和音乐不能很好的融合，就会严重影响演唱者的唱感，会产生跑调，跟不上节拍等各种现象。
二十世纪欧美流行乐36位最伟大的女歌手1）Ella Fitzgerald“艾拉费兹杰拉”——1918年生于弗吉尼亚州，1994年去世。50-60年代第一天后，美国百老汇“急智歌后”，被誉为爵士第一夫人。歌声如水晶般清澈的她，音色中充满了女性万般风情的魅力。她的幽默心声都是近半世纪美式风情的诠释者。1994年去世。 代表作《misty blue》、《louis armstrong》 2）Aretha Franklin“艾瑞莎弗兰克林”——美国灵魂乐的巨人、Pop界的巨星。以烈焰般燃烧的气质征服了无数听众。她将灵乐虔诚的信仰表现得淋漓尽致，并跨越流行与黑人音乐界线。作为第一位女性艺人进入摇滚名人殿堂。 代表作《i never loved a man before》、《i say a little prayer》、《chain of fools》、《a rose is still a rose》、《Respect》 3）Anne Murray“安妮莫莉”——全球著名的民谣摇滚巨星。被认为是为加拿大通向国际成功的女性先驱！Anne走民谣摇滚的路线，她将节奏布鲁斯和主流流行乐的风格兼容并蓄在自己的歌曲中，叱咤流行乐坛。 代表作《you needed me》、《i just fall in love again》、《you light up my life》、《dream a little dream of me》、《broken hearted me》 4）Annie Lennox“安妮兰妮克丝”——80、90年代红遍欧美、蜚声国际流行乐坛20年，有“首席女伶”美誉！1981年组建Eurythmics乐队，成为最受欢迎的新浪潮乐队。90年代她以磅礴大气的半歌剧式演唱，空灵、深邃犹如天籁般的优美歌声直触心灵，被称为现代最优美女声之一。
?vega vegas s视频制作大型图文教程(十九节初级内容完结,没什么人气...就到此腰斩吧...end 前言 玩VEGAS也有半年了,也学会了一点皮毛。今天就做一个图文教程送给NGA 玩家众了。此教程从零开始和大家一起学习制作魔兽世界视频,我尽量用大家都看得懂的表达方式....写的错误的地方希望大家纠正。另外不管是魔兽世界还是制作视频,他们仅仅只是一个游戏,虽然有这么一点小小的成就感,但是请不要忘“本”我将会一直更新直到写完为止... 初级篇 一.对于vegas的定位 长篇文字什么的我就不Ctrl+C Ctrl+V了,简单地说就是一个视频制作软件, 优点就是制作上手快,剪辑灵活,特效美观,当然了再完美的东西还是有瑕疵,缺点就是合成部分不如AE,原因就是VEGAS仅仅只是视频制作软件,而AE是一个 影视特效合成软件,最后一点比较遗憾的是无法直接导入PS和RP的工程项目....看完之后大家也不要太灰心,我们仅仅只是针对制作游戏视频,用Vegas绰绰有余
二.VEGAS的下载及安装 电驴下载地址[ https://www.doczj.com/doc/3618712759.html,/topics/2718912/ 此网页不属于本网站,不保证其安全性 继续访问取消不再提示我 https://www.doczj.com/doc/3618712759.html,/topics/2718912/] 大家可以参照这个帖子进行安装 [https://www.doczj.com/doc/3618712759.html,/read.php?tid=3284424&_fp=1]另外通过一点好友的回复,我发现安装还是存在一定的误区,在所有的步 骤最前面我推荐大家用360软件管理下载C+2005DX9.0C+2008然后开始安装(PS :安装教程在/Vegas.Pro.v8.0c智能安装/v8.0/安装全过程——这个东西下 载后自带。最后说一下WIN764位的我不太清楚,本人系统为XP-SP3。
效果器使用原理 效果器对于录音来说，就象是你烹饪时所加入的香料--它们可以非常有效地增强现有声音的感染力，但是要想使用好这些效果器，你必须要经过一个漫长的学习过程。遗憾的是有很多人对他们的效果器非常陌生，在使用时通常都是随意地设一个值，然后就异想天开地指望得到精彩的声音。 如果你知道了这些方盒子是如何进行工作的，你就可以更加有效地使用它们。在下面的文章中，我们不仅列出了一些效果器通常的使用规则，还向你讲述了它们的一些重要参数、经常给我们带来麻烦的地方以及一些应用热点。 压限器（Compressor/Limiter） －概述 压缩器/限制器（compressor/limiter，简称压限器）的用途是让信号的输出动态范围变小，它使较微弱的信号变大而使较大的信号变小。其结果是使大信号与小信号之间的差别变小。例如，压限器可以用来使snare鼓的音轨变得平淡柔和，允许整个鼓的声音在混音器上被提升到一个较高的电平，而不会使母带过载。对于有些歌手来说，他们在进行录音时总是不能够很好地保持嘴部与麦克风之间的距离，这时候使用一些温和的压缩效果就可以使得人声音轨的表现更佳。 －工作原理 一旦输入的信号电平超过了用户设定的阀值，则压缩器就将开始工作，把过高的输入电平降低。这样得到的结果是，在增大输入电平的同时，不会造成输出电平产生同等幅度的增大。例如，设置压缩率为2：1，则每增加2 dB的输入电平只会造成输出电平有1 dB的变化。 －重要的参数 阀值（threshold）参数：决定了要被压缩或是限制的信号的上下限。处于阀值以内的信号将不会受到影响。 比率（ratio）参数：选择了在输入信号超过阀值时，输出电平改变的方式。较高的比率值，将导致较大的压缩，并使得声音听起来很"挤"。非常高的比率值会导致信号产生极端的"上限成分"（ceiling）。这叫做极限（limiting）。 输出（output）参数：提高增益可以抵消掉由于动态范围约束而产生的较低的电平。
调音台效果器等设备的 使用方法 Document serial number【UU89WT-UU98YT-UU8CB-UUUT-UUT108】
效果器的连接和应用效果器的连接和应用 一、效果器在系统中的连接： 1、串行连接 乐器或者乐器话筒直接输入效果器输入接口，效果器处理后通过输出接口连接调音台LINE接口。或者在调音台INSERT断点接口串接也行。 2、发送返回连接 调台辅助输出(AUX OUT)，连接到效果器输入接口，效果器处理后通过输出接口返回调音台LINE IN线路输入通道或者AUX RETURN辅助返回通道。 二、效果的分类 1、延时效果（或称为者回声效果） 延时效果源于山歌，歌声在山谷中回荡，经过大山的反射出现多次的回声，根据回声与原始声的时延我们称为延时效果(DELAY)或者称为回声效果（ECHO）,另外环境不同产生的回声次数不同，听觉的效果也不同，根据回声次数我们称为延时深度（DELAY DEPTH）或者回声数（REPEAT）。 延时效果可以掩饰演唱者的气息和连贯性不好等缺陷，但使用这种效果会降低演唱者的清晰度，一般对业余歌手或KTV中使用 2、混响效果（或称为残响效果） 混响效果源于西方的歌剧院，通过建筑物内的多次反射产生的混响声效果（REVERB）。混响声持续的时间就是混响时间（REV TIME）,一般建筑声学把混响声衰减60dB的时间定义为混响时间。另外原始声到第一次出现的混响效果声
的时间，效果器上称为预延时时间（PRE DELAY）。为了切除多余的尾音或者说缩短混响时间通过衰退时间(DECAY)来调节。 混响效果是我们比较常用的效果类型，要设置的好就需要不断实践和总结。 三、结合实物理解一下应用 我们拿TC的M350来理解吧，因为这个效果器包括了延时效果和混响效果两个引擎。我只是想结合实物了解一下调节键的情况，方便理解。 我们主要分主控部分，延时效果部分，混响效果部分 1、主控部分 INPUT GAIN：输入效果器的增益调节 MIX RATIO：干湿混合比调节。是调台辅助输出然后返回调台的接发，这个要打到最大，让效果器输出的完全是效果声（湿信号），我们在调台上调节原始声和效果声的比例。 EFFFECT BAL：延时效果和混响效果的比例调节。12点位置两效果引擎均有最大输出，顺时针或逆时针打到最大，意味着其中一个效果引擎旁通。 2、延时效果部分： DELAY EFFECTS：延时效果模式的选择，共15种效果模式，打到OFF为关闭 DELAY/TIMING：调节延时时间，就是原始声和效果声之间的时间间隔。或者说回声的间隔时间。 FEEDBACK/DEPTH：延时深度调节，就是回声的次数，一般调到3-5次回声即可。 注：在选择完延时效果模式后，延时时间和延时深度的调节旋转在12点位置是厂家推荐的设置，我们可以根据时间情况进行微调即可。