Efficient Nash computation in large population games with bounded influence

Michael Kearns

Department of Computer and Information Science University of Pennsylvania

Philadelphia,Pennsylvania

mkearns@https://www.wendangku.net/doc/d710553478.html,

Yishay Mansour School of Computer Science Tel Aviv University

Tel Aviv,Israel

mansour@cs.tau.ac.il

Abstract

We introduce a general representation of large-

population games in which each player’s in?u-

ence on the others is centralized and limited,but

may otherwise be arbitrary.This representation

signi?cantly generalizes the class known as con-

gestion games in a natural way.Our main results

are provably correct and ef?cient algorithms for

computing and learning approximate Nash equi-

libria in this general framework.

1INTRODUCTION

We introduce a compact representation for single-stage ma-trix games with many players.In this representation,each player is in?uenced by the actions of all others,but only through a global summarization function.Each player’s payoff is an arbitrary function of their own action and the value of the summarization function,which is determined by the population joint play.This representation of large-population games may be exponentially more succinct that the naive matrix form,and here we prove that vast compu-tational savings can be realized as well.A natural example of such games is voting scenarios(a special case in which the summarization function is both linear and symmetric), where each player’s payoff depends only on their own vote and the outcome of the popular election,but not on the de-tails of exactly how every citizen voted.As discussed be-low,summarization games generalize a number of existing representations in the game theory literature,such as con-gestion games.

We make the natural assumption of bounded in?uence—that is,that no player can unilaterally induce arbitrarily large changes in the value of the summarization function. (V oting is a simple example of bounded in?uence.)Un-der only this assumption and a bound on the derivatives of the private individual payoff functions(both of which ap-pear to be necessary),we give an algorithm for ef?ciently computing approximate Nash equilibria,which interest-ingly always outputs pure approximate equilibria.We also prove that a simple variant of distributed smoothed best-response dynamics will quickly converge to(learn)an ap-proximate equilibrium for any linear summarization func-tion.These algorithms run in time polynomial in the num-ber of players and the approximation quality parameter,and are among the few examples of provably ef?cient Nash computation and learning algorithms for broad classes of large-population games.

2RELATED WORK

A closely related body of work is the literature on games known as congestion games(Rosenthal[1973])or exact potential games(Monderer and Shapley[1996]),which are known to be equivalent.In congestion games and their gen-eralizations,players compete for a central resource or re-sources,and each player’s payoff is a(decreasing)function of the number of players selecting the resources.An ex-ample is the well-studied Santa Fe Bar problem,where pa-trons of a local bar receive positive payoff if their numbers are suf?ciently low,negative payoff if they exceed capac-ity,and players who stay home receive0payoff.Single-resource congestion games can be viewed as summariza-tion games in which the global summarization is symmet-ric—that is,dependent only on the total number of play-ers selecting the resource.In the current work we allow the summarization function to be both non-symmetric and non-linear,but our results can also be viewed as a contribution to the congestion game literature.While a fair amount is understood about the mathematical properties of equilibria in congestion games(such as the existence of pure equilib-ria),and there has been a great deal of recent experimen-tal simulation(see,for example,Greenwald et al.[2001]), there seems to be relatively little work providing provably ef?cient and correct algorithms for computing and learning equilibria.

We also view the proposed representation and algorithms as complementary to recent work on compact undirected graphical models for multi-player games(Kearns et al. [2001],Littman et al.[2002],Vickrey and Koller[2002]). While those works emphasize large-population games in

which each player is strongly in?uenced by a small num-

ber of others,the current work focuses on games in which each player is weakly in?uenced by all others.This is

analogous to the two main cases of tractable inference in

Bayesian networks,where the polytree algorithm provides an algorithm for sparse networks,and variational algo-

rithms yield approximate inference in dense networks with small-in?uence parametric CPTs.

3DEFINITIONS AND NOTATION

We begin with the standard de?nitions for multiplayer ma-

trix games.An-player,two-action game is de?ned by

a set of matrices(),each with in-dices.The entry speci?es the payoff to player when the joint action of the players is

Thus,each has entries.We shall assume all payoffs are bounded in absolute value by1.

The actions0and1are the pure strategies of each player, while a mixed strategy for player is given by the proba-bility that the player will play0.For any joint mixed strategy,given by a product distribution,we de?ne the expected payoff to player as, where indicates that each is0with probability and1with probability.

We use to denote the vector which is the same as except in the th component,where the value has been changed to.A Nash equilibrium for the game is a mixed strategy such that for any player,and for any value ,.(We say that is a best response to the rest of.)In other words,no player can im-prove their expected payoff by deviating unilaterally from a Nash equilibrium.The classic theorem of Nash[1951] states that for any game,there exists a Nash equilibrium in the space of joint mixed strategies.

We will also use a straightforward de?nition for approxi-mate Nash equilibria.An-Nash equilibrium is a mixed strategy such that for any player,and for any value ,.(We say that is an-best response to the rest of.)Thus,no player can improve their expected payoff by more than by deviating unilaterally from an approximate Nash equilibrium.

As in Kearns et al.[2001],our goal is to introduce a nat-

ural new representation for large multiplayer games that is considerably more compact than the classical tabular form, which grows exponentially in the number of players.How-ever,rather than succinctly capturing games in which each player has a small number of possibly strong“local”in?u-ences,our interest here is at the other extreme—where each player is in?uenced by all of the others in a large pop-ulation,but no single player has a dominant in?uence on any other.

there are players,in many natural bounded-in?uence set-

tings the maximum in?uence will be on the order of(as in voting),or at least some function diminishing with. Finally,we discuss the individual payoff functions.Here we simply assume that each player possesses separate

payoff functions for their two actions,and.If the

pure actions of the other players are given by the vec-tor(where here is irrelevant)the payoff to for playing is de?ned to be,and the payoff to player for playing is de?ned to be.Thus,for any joint play,each player is told what values their two

actions will yield for the population summarization func-tion,and has private payoff functions indicating their own reward for each resulting value.We assume that the are real-valued functions mapping to.We shall also assume that all the payoff functions are continuous and have bounded derivatives.

Note that even though the summarization function has

bounded in?uence,and thus a player’s action can have only

bounded effect on the payoffs to others,it can have dra-matic effect on his own payoffs,since and may as-sume quite different values for any mixed strategy(despite the bounds on their derivatives).We feel this is a natural model for many large-population settings,where subjective (private)regret over actions may be unrelated to the(small) in?uence an individual has over global quantities.For in-stance,a staunch democrat might personally?nd voting for a republican candidate quite distasteful,even though this individual action might have little or no effect on the over-all election.It is their private payoff functions that makes individuals“care”about their actions in a large population where the global effects of any single player are negligible. We shall assume throughout that the summarization func-tion and all the payoff functions can be ef?ciently computed on any given input;formally,we will assume such a computation takes unit time.Thus,the tuple

,which we shall call a(large popu-lation)summarization game,is a representation of an-player game that may be considerably more compact than its generic matrix representation.We say that is a-summarization game if the in?uence of is bounded by, and the derivatives of all are bounded by.

Two?nal remarks on the de?nition of a summarization game are in order.First,note that the representation is entirely general:by making the summarization and pay-off functions suf?ciently complex,any-player game can be represented.If outputs enough information to recon-struct its input(for example,by computing a weighted sum of its inputs,where the weight of bit is),and the payoff functions simply interpolate the values of the orig-inal game matrices for player,the original game is ex-actly represented.However,in such cases we will not have small in?uence and derivatives,and our results will natu-rally degrade.It is only for bounded in?uence and deriva-tive games,which seem to have wide applicability,that our results are interesting.Second,we note that if we view the summarization function as being de?ned for every in-put length(as in voting)and?xed,and the continuous payoff functions as being?xed,then summarization games naturally represent games with an arbitrarily large or grow-ing number of players,and our results will shed light on computing and learning equilibria in the limit of large pop-ulations.

The results in this paper describe ef?cient algorithms for computing and learning approximate Nash equilibria in summarization games,and provide rates of convergence as a function of summarization in?uence,payoff derivatives, and population size.We now turn to the technical develop-ment.

4COMPUTING EQUILIBRIA IN

SUMMARIZATION GAMES

The?rst of our two main results is an ef?cient algorithm for computing approximate Nash equilibria in bounded-in?uence summarization games:

Theorem1There is an algorithm SummNash that takes as input any and any-summarization game

over players,and outputs an -Nash equilibrium for.Furthermore,this ap-proximate equilibrium will be a pure(deterministic)joint strategy.The running time of SummNash is polynomial in ,,and.

Before presenting the proof,let us brie?y interpret the re-sult.First,the accuracy parameter is an input to the al-gorithm,and thus can be made arbitrarily small at the ex-pense of the polynomial dependence on of the running time.As for the term in the approximation quality,it is natural to think of the derivative bound as being a?xed constant,while the in?uence bound is some diminishing function of the number of players—that is,individuals have relatively smooth payoffs independent of population size,while their individual in?uence on the summarization function shrinks as the population grows.Under such cir-cumstances,Theorem1yields an algorithm that will com-pute arbitrarily good approximations to equilibria as the population increases.

The proof of Theorem1and the associated algorithm will be developed in a series of lemmas.Our?rst step is to approximate the continuous,bounded-derivative indi-vidual payoff functions by piecewise constant(step) functions.For a given resolution(to be deter-mined by the analysis),we divide into the-intervals

.Denote the th such interval as.We de?ne the approximation

to be constant over any-interval.Speci?cally,for any

,.Since the derivative of the is bounded by,we have for all play-ers,,and.In the sequel,we shall refer

to as the-approximate summa-rization game for.

We?rst show that the bounded derivatives of the payoff functions translates to a Lipschitz-like condition on the ap-proximate payoff functions.

Lemma2For all,

where we de?ne if and otherwise.

Proof:Clearly the difference is0if.If we have

where the?rst inequality comes from the approximation quality,and the second from the bound on the derivatives of the.

The following straightforward lemma translates the quality of approximate Nash equilibria in back to the original game.

Lemma3Let be any-Nash equilibrium for the-approximate summarization game.Then is a-Nash equilibrium for.

Proof:Since is a-Nash equilibrium for,each player is playing a-best response.The rewards in can change by at most for each action,which implies that the change to a new best response is at most.

We next give a lemma that will simplify our arguments by letting us de?ne(approximate)best responses solely in terms of the single global summarization value,rather the multiple local values for each and.We start with the following de?nition.

De?nition1Let be the-approximate summarization game,and let be any mixed strategy.We de?ne for player the single-value apparent best response in as

Thus,is the apparent best response for in to if ignores the effect of his own actions on the summarization function.We now show that this apparent best response in is in fact an approximate best response in.

Lemma4Let be the-approximate summarization game.Let be any mixed strategy.Then is a -best response for player to in.Proof:For any pure strategy and any,we have

due to the bound on in?uence.Note that the right-hand side of this inequality is if and0if.By Lemma2,we have

Taking expectations under gives

Now if

(that is,is not already a true best response to for in ),then the inequality above implies it is a-best response.

Now note that by construction,if is any-interval,and and are any two pure strategies such that

and(that is,both vectors give a value of the summarization function lying in the same-interval),then

,because the approximate payoff functions do not change over.Furthermore,the action

is an approximate best response for in by Lemma4.In other words,in,we have reduced to a setting in which the (approximate)best response of all players can be viewed solely as a function of the-interval in which lies, and not on the details of itself.

For any-interval,let us de?ne

where is any vector such that.Thus,

is the vector of(apparent)best responses of the players in when the value of the summarization function falls in. This best response itself gives a value to the summarization function,which we de?ne as.We can extend this de?nition to view as a mapping from

to(rather than from-intervals to)by de?ning to be,where.In Figure1we provide a sample plot of a hypothetical that we shall refer to for expository purposes.

The purpose of the de?nition of is made clear by the next lemmas.The intuition is that those places where “crosses”the diagonal line are indicative of approx-imate Nash equilibria.We begin with the easier case in which crosses the diagonal during one of its constant-valued horizontal segments,marked as the point in Fig-ure1.

Lemma5Let be an-interval such that. Then is a-Nash equilibrium for.

Proof:Let.Since,every player is playing,and thus by Lemma4,a -best response to.

We next examine the case where Lemma5does not apply. First we establish a property of the function.

Lemma6If for every-interval,,then there exists a such that. Proof:For we have,and for

(the last interval)we have.Therefore there has to be a for which the lemma holds.

In other words,if Lemma5does not apply,there must be two consecutive intervals whose-values“drop”across the diagonal.This case is illustrated in Figure1by the vertical dashed line containing the point.

Lemma7Let be such that. Then there is a pure strategy which is a-Nash equilibrium in.

Proof:Let and.Let be the number of indices for which.De?ne the sequence of vectors such that,,and for every,is obtained from by?ipping the next bit such.Thus,in each,bits that have the same value in and are unaltered,while bits that differ in and?ip exactly once during the“walk”from to. The intuition is that if we can?nd any vector on the walk which gives a value to falling in or near the interval, it must be an approximate Nash equilibrium,since players whose best response in this neighborhood of-values may be strongly determined(that is,those for which) are properly set throughout the walk,while the others may be set to either value(since they are nearly indifferent in this neighborhood,as evidenced by their switching approx-imate best response values from to).In Figure1, the points along the vertical line that include conceptu-ally show the different values of during the walk. Now for each we de?ne.Due to the bounded in?uence of,we have that for all .This implies that for some value,we have

.(In Figure1,point corresponds to the vector on the walk whose-value comes closest to,and thus constitutes the desired approximate Nash equilibrium.) We now show that is an approximate Nash equilibrium. Consider player,and assume that,but that the best response for is actually(the other cases are similar).If ,then is a-best response by Lemma4. Otherwise,.Since is the apparent best response for,we have

1

1

1

1

1

1

1

1

1

1

1

1

1

1

A

B

Y = X

Y = V(I)

Y

X = I

Figure1:Hypothetical plot of the function over the-intervals .We view the-axis as being both a continuum of individual points and a discrete sequence of-intervals.generally begins above the diagonal line,and ends up below the diagonal, and thus must cross the diagonal at some point.The point labeled is an example of a horizontal crossing as covered by Lemma5. The column of points including the point labeled is an exam-ple of a vertical crossing as covered by Lemma6.Each point in this column indicates a value of realized on the vector“walk”discussed in the proof of Lemma6,while the point itself is the value of nearest the diagonal in this walk.The analysis estab-lishes that(at least)one of these two crossing cases must occur.

We now bound the difference in the reward to player due to playing action rather than in response to,which is

This difference is bounded by

since,and by Lemma2.Again by Lemma2, we have

Putting it all together gives

Therefore every player is playing a-best re-sponse.

Lemma5covers the case where crosses the diagonal hor-izontally,and Lemma7the complementary case where it does not.Either case leads to the ef?cient computation of a -Nash equilibrium for,which by Lemma3 is a-Nash equilibrium for the original game .Setting or completes the proof of Theorem1.

A description of the algorithm SummNash detailed in the above analysis is provided in Figure2.

Algorithm SummNash

Inputs:An-player-summarization game,and an approximation parame-

ter.We assume that and each payoff function are provided as black-box subroutines that cost unit

computation time to call.

Output:A-Nash equilibrium for.

1..

2.For each player,and,construct the-approximate payoff functions:for

every-interval,for all.Note that this requires

evaluations of each.

3.Construct the mapping by setting for every-interval.

4.For each-interval,,check if.If so,then output.

5.For each-interval,,check if.If so:

(a)Let be the set of player indices whose play is the same in and.

(b).

(c).

(d).

(e)For,let be obtained from by?ipping the bit corresponding to the next index

in.Note that.

(f)For,if,output.

Figure2:Algorithm SummNash.

5LEARNING IN LINEAR

SUMMARIZATION GAMES

In this section,we propose and analyze a distributed learn-

ing algorithm for the players of a game in which the sum-

marization function is linear,that is,.

It is easily veri?ed that in such a case that each in?uence

.Our main result is that the learning algorithm con-

verges to an approximate Nash equilibrium for the summa-

rization game.The analysis will rely heavily on the tools

developed in Section4.

The learning algorithm for each player is rather simple,

and can be viewed as a variant of smoothed best response

dynamics(Fudenberg and Levine[1999]).First of all,if

is the original(linear)summariza-

tion game,each player will instead use the-approximate

payoff functions described in Section4.Note that these

approximations can be computed privately by each player.

We shall use to denote the joint mixed strategy of the

population at time in the learning algorithm.In our learn-

ing model,at each round,the expected value of the sum-

marization function

is broadcast to all players.

.Thus,as before we have improved

approximated with larger populations.Note,however,that

the Pigeonhole Principle.

Lemma11Assume that the mean never visits any-interval twice.Then SummLearn()converges after at most steps.

The following lemma handles both the case that learning never converges,and the case that some interval is visited twice.

Lemma12Suppose that at some time,makes a sec-ond visit to some-interval.Then for all,the mixed strategies are all-Nash equilibria for,where. Proof:Let be the?rst-interval visited twice by, and that is the?rst step of the second visit.Since, at time we had either or; we assume the latter case without loss of generality.Note that that since is the?rst revisited-interval,is mono-tonically increasing while,and monotonically de-creasing while.Thus,for all,we will have.

Consider a player such that.After at most time steps we will have

.Therefore,as in the proof of Lemma9,by time player will play an-best response in for all

.For the other players such that,any action is an-best response for all, since.

For the case,we thus have the following theorem, which together with Theorem15below constitutes the sec-ond of our main results.

Theorem13After at most steps, SummLearn()plays an-Nash equi-librium for all subsequent time steps.

As in algorithm,We can make the term smaller than any desired by choosing

and,with the resulting polynomial dependence on in the running time.This leaves us with the un-controllable term.Again,as we have discussed,in many reasonable large-population games we expect these in?uence terms to vanish as becomes large.Also,note that given that we require the learning rate,there is no bene?t to setting much smaller than this,and thus the choice yields an overall convergence time of

.

We now analyze SummLearn().Here we cannot expect to upper bound the time it will take to converge to an ap-proximate Nash equilibrium—technically,if the interval is such that,might stay in for steps(see Figure3).Since can be arbitrary small,this time cannot be bounded.However,we can show that any time is“near the diagonal”,the players are play-ing an approximate Nash equilibrium.This implies that we can bound the number of time steps in which they are not playing an approximate Nash equilibrium.

Lemma14Consider a visit of to interval in the time interval.For all times

is an-Nash equilibrium.

Proof:As in Lemma9,after steps we have

.Therefore each player is playing a

-best response in.

Since Lemma12does not depend on the termination mech-anism,Lemma14and the choice implies the fol-lowing theorem.

Theorem15For any,if SummLearn()is run for an in?nite sequence of steps,the players play an

-Nash equilibrium in all but at most

steps.

Note that though SummLearn()has no dependence on (only the global summarization mean must be broadcast), Theorem15provides a spectrum of statements about this algorithm parameterized by—as we reduce,we give worse(larger)bounds on the total number of steps that a better approximation to equilibrium is played.

References

D.Fudenberg and D.Levine.The Theory of Learning in Games.

MIT Press,1999.

A.Greenwald,E.Friedman,and S.Shenker.Learning in network

contexts:Results from experimental simulations.Games and Economic Behavior:Special Issue on Economics and Arti?cial Intelligence,35(1/2):80–123,2001.

M.Kearns,M.Littman,and S.Singh.Graphical models for game theory.In Proceedings of UAI,2001.

M.Littman,M.Kearns,and S.Singh.An ef?cient exact algorithm for singly connected graphical games.In Neural Information Processing Systems,2002.To appear.

D.Monderer and L.Shapley.Potential games.Games and Eco-

nomic Behavior,14:124–143,1996.

J.F.Nash.Non-cooperative games.Annals of Mathematics,54: 286–295,1951.

R.Rosenthal.A class of games possessing pure-strategy Nash equilibria.International Journal of Game Theory,2:65–67, 1973.

D.Vickrey and D.Koller.Multi-agent algorithms for solving

graphical games.In Proceedings of the National Conference on Arti?cial Intelligence(AAAI),2002.To appear.

kano模型

kano模型 目录 1简介 2内容分析 3需求分析 4操作意义 5优缺点 6满意度 7质量划分 8有关评价 简介 受行为科学家赫兹伯格的双因素理论的启发，东京理工大学教授狩野纪昭(Noriaki Kano)和他的同事Fumio Takahashi于1979年10月发表了《质量的保健因素和激励因素》(Motivator and Hygiene Factor in Quality)一文，第一次将满意与不满意标准引入质量管理领域，并于1982年日本质量管理大会第12届年会上宣读了《魅力质量与必备质量》﹙Attractive Quality and Must-be Quality﹚的研究报告。该论文于1984 年1月18日正式发表在日本质量管理学会(JSQC)的杂志《质量》总第l4期上，标志着狩野模式(Kano mode1)的确立和魅力质量理论的成熟。 2内容分析编辑

KANO模型定义了三个层次的顾客需求:基本型需求、期望型需求和兴奋型需求。这三种需求根据绩效指标分类就是基本因素、绩效因素和激励因素。 基本型需求是顾客认为产品“必须有”的属性或功能。当其特性不充足（不满足顾客需求）时，顾客很不满意；当其特性充足（满足顾客需求）时，无所谓满意不满意，顾客充其量是满意。 期望型需求要求提供的产品或服务比较优秀，但并不是“必须”的产品属性或服务行为有些期望型需求连顾客都不太清楚，但是是他们希望得到的。在市场调查中，顾客谈论的通常是期望型需求，期望型需求在产品中实现的越多，顾客就越满意;当没有满意这些需求时，顾客就不满意。 兴奋型需求要求提供给顾客一些完全出乎意料的产品属性或服务行为，使顾客产生惊喜。当其特性不充足时，并且是无关紧要的特性，则顾客无所谓，当产品提供了这类需求中的服务时，顾客就会对产品非常满意，从而提高顾客的忠诚度。 3需求分析 基本品质（需求） kano模型 也叫理所当然品质。如果此类需求没有得到满足或表现欠佳，客户的不满情绪会急剧增加，并且此类需求得到满足后，可以消除客户的不满，但并不能带来客户满意度的增加。产品的基本需求往往属于此类。对于这类需求，企业的做法应该是注重不要在这方面失分。 期望品质（需求） 也叫一元品质。此类需求得到满足或表现良好的话，客户满意度会显著增加，当此类需求得不到满足或表现不好的话，客户的不满也会显著增加。这是处于成长期的需求，客户、竞争对手和企业自身都关注的需求，也是体现竞争能力的需求。对于这类需求，企业的做法应该是注重提高这方面的质量，要力争超过竞争对手。魅力品质（需求） 此类需求一经满足，即使表现并不完善，也能到来客户满意度的急剧提高，同时此类需求如果得不到满足，往往不会带来客户的不满。这类需求往往是代表顾客的潜在需求，企业的做法就是去寻找发掘这样的需求，领先对手。

卡诺模型

. 卡诺模型 年提和他的同事FumioTakahashi1984于日本东京理工大学教授狩野纪昭(Noriaki Kano)。二维包含了两个维(Two-dimension Model)(Kano Mode]),又称作二维品质模型出了卡诺模型属,;从用户对产品的满意度进行考量度:从产品的品质角度考虑,属于客观的产品机能或功能于用户的主观感受。一维品质重要理论模型。其中的品质主要包括个四部分: 卡诺模型是产品品质创造。一(Attractive)、无差异品质(Indifference)(One-dimensional)、必要品质(Must-be)、魅力品质用户满意也因之提升；如不提供此维品质又称作线性品质、期望品质：当需求越得到满足,则会感到不满。一般而言品质越好，满意度越高，反之则受到负面评价；因此满意度需求, 与品质成正比。以获得更好的用户要求设计师聚焦在核心需求及其体验的优化,在设计策略中,一维品质满意度。产品的功能性、可用性、易用性及可扩展性都可以对一维品质造成影响。用户满意当需求得到优化时,必要品质是产品的基本要求。由于用户的满意度会有上限,必要品质要求设在设计策略中,;当不提供此需求时,用户满意度会大幅降低。度不一定会提升尽可能地满足用户的所有需求。因此设计策略要通过分,计师进行严谨而又细致的统筹工作析用户需求定义明确的产品功能。用户满意度不会,魅力品质是一种用户意想不到的品质。若不提供用户意想不到的需求魅力增幅远高于一维品质。,在设计策略中,降低;而若提供此需求,用户满意度则会有较大提升但往往成为品质是对产品创新及创新优良体验的追求。它在设计中不涵盖产品的所有模块,魅力品质需要建立产品的点睛之笔。每一个创新优良的体验都能为产品增加魅力值。因此,以发掘真正具有价值的品通过挖掘他们潜在的需求寻找设计的创新点,在目标用户的基础上, 质。不与用户满意度关联。无论,无差

KANO模型详解

最早在腾讯的《在你身边为你设计》中看到这个模型，却一直没完全弄懂是怎么使用的，今天自己编造了一些数据，一步步做了一遍，总算理解了。 以下的引用部分引用自知乎。 1.卡诺模型简介-对用户满意度和需求进行分析的工具卡诺模型（KANC模型）是对用户需求分类和优先排序的有用工具，以分析用户需求对用户满意的影响为基础，体现了产品性能和用户满意之间的非线性关系。在卡诺模型中，将产品和服务的质量特性分为四种类型：⑴必备属性；⑵期望属性；⑶魅力属性；⑷无差异属性。 满意SiBi A 满意度低 KANO模型中的几种属性魅力属性：用户意想不到的，如果不提供此需求，用户满意度不会降低，但当提供此需求，用户满意度会有很大提升； 期望属性：当提供此需求，用户满意度会提升，当不提供此需求，用户满意度会降低； 必备属性：当优化此需求，用户满意度不会提升，当不提供此需求，用户满意度会大幅降低； 无差异属性：无论提供或不提供此需求，用户满意度都不会有改变，用户根本不在意； 反向属性：用户根本都没有此需求，提供后用户满意度反而会下降2.KANO模型的使用-问卷编制与数据处理 KANO问卷对每个质量特性都由正向和负向两个问题构成，分别测量用户在面对存在或不存在某项质量特性时的反应。需要注意： ①KANO可卷中与每个功能点相关的题目都有正反两个问题，正反问题之间的区别需注意强调，防止用户看错题意； ②功能的解释：简单描述该功能点，确保用户理解；

③选项说明：由于用户对“我很喜欢”“理应如此”“无所谓”“勉强接受” “我很不喜欢” 的理解不尽相同，因此需要在问卷填写前给出统一解释说明，让用户有一个相对一致的标准，方便填答。 我很喜欢：让你感到满意、开心、惊喜。 它理应如此：你觉得是应该的、必备的功能/ 服务。 无所谓：你不会特别在意，但还可以接受。 勉强接受：你不喜欢，但是可以接受。 我很不喜欢：让你感到不满意。 因此在编制问卷的时候，对每个项目都要有正反两道题来测，比如，“如果在中加入朋友圈功能，您怎样评价？”对应“如果在中去掉朋友圈功能，您怎样评价？” 均提供五个选项：我很喜欢、它理应如此、无所谓、勉强接受、我很不喜欢 那么每个用户对于某一个项目的态度必然落入下图表中的某个格子。而对所有的用户来说，共有5*5 即25种可能，统计每种可能下的用户人数占总人数的百分比，来填入下表。之后将下表中标A、O Ml、R、Q的格子中百分比相加，即可得到五种属性对应的百分比。从需求的角度来说，先满足M百分比最高的去掉R百分比最高的，再满足O百分比最高的，最后满足A百分比最高的。

Kano模型的数据统计分析

Kano模型的数据统计分析 1、用户需求分类 1．1Kano模型 可以把基本品质、期望品质、和魅力品质理解为客户对产品的要求：功能要求---性价比/品牌效应---附加值/特殊性。 1．2用户需求分类 将每项用户需求按照Kano模型进行分类，即分为基本品质、期望品质和惊喜品质。先进行用户意见调查，然后对调查结果进行分类和统计。 1．2．1市场调查 对每项用户需求，调查表列出正反2个问题。例如，用户需求为“一键通紧急呼叫”，调查问题为“一键通紧急呼叫能随呼随通，您的感受如何?”以及“一键通紧急呼叫不能随呼随通，您的感受如何?”，每个问题的选项为5个，即满足、必须这样、保持中立、可以忍受和不满足。

注：√表示用户意见 1．2．2调查结果分类 通过用户对正反2个问题的回答，分析后可以归纳出用户的意见。例如，对某项用户需求，用户对正向问题的回答为“满足”，对反向问题的回答为“不满足”，则用户认为该项需求为“期望品质”。每项用户需求共5×5—25个可能结果。 基本品质、期望品质和惊喜品质是3种需要的结果。其他3种结果分别为可疑、反向和不关心，这是不需要的，必须排除。 (1)可疑结果（用户的回答自相矛盾）。可疑结果共2个，即用户对正反问题的回答均为“满足”或“不满足”。例如，对于“一键通紧急呼叫”，正向

问题为“一键通紧急呼叫能随呼随通，您的感受如何?”，用户回答是“满足”；反向问题为“一键通紧急呼叫不能随呼随通，您的感受如何?”，用户回答还是“满足”。这表明无论一键通紧急呼叫是否能随呼随通，用户都会满足，这显 然是自相矛盾的。出现可疑结果有2种可能：一是用户曲解了正反问题，二是用户填写时出现错误。统计时需要去除可疑结果。 (2)反向结果（用户回答与调查表设计者的意见相反）。正向问题表明产品 具有某项用户需求，反向问题表明不具备该用户需求，正向问题比反向问题具 有更高的用户满意，但用户回答却表明反向问题比正向问题具有更高的客户满 意度。例如，对用户需求“一键通紧急呼叫”，正向问题为“一键通紧急呼叫 能随呼随通，您的感受如何?”，用户回答为“不满足”，反向问题为“一键通紧急呼叫不能随呼随通，您的感受如何?”，用户的回答为“满足”，这显然与调查表设计者的意见相反。反向结果较多时，表明调查表的设计存在问题，需 要改进。 (3)不关心（用户对调查表所提出的问题漠不关心）。例如，对用户需求 “一键通紧急呼叫”，正向问题为“一键通紧急呼叫能随呼随通，您的感受如何?”，用户回答为“保持中立”，反向问题为“一键通紧急呼叫不能随呼随通，您的感受如何?”，用户回答还是“保持中立”，说明用户对“一键通紧急呼叫” 的存在与否，既不是满足，也不是不满足。统计时需要去除这类结果。 1．2．3调查结果统计 调查用户意见后，需要通过统计分析来判断每项用户需求属于哪种品质。 判定方法是：对调查结果进行统计，去除可疑、反向和不关心结果，根据基本、期望和惊喜3种品质统计结果的数量，确定用户需求属于哪种品质。例如，对用户需求“一键通紧急呼叫”，如通过统计调查结果表明，用户认为“一键通紧 急呼叫”是“基本品质”的最多，那么用户需求“一键通紧急呼叫”被确定为 基本品质。

kano模型的详尽解释

kano模型的详尽解释 受行为科学家赫兹伯格的双因素理论的启发，东京理工大学教授狩野纪昭(Noriaki Kano)和他的同事Fumio Takahashi 于1979年10月发表了《质量的保健因素和激励因素》(Motivator and Hygiene Factor in Quality)一文，第一次将满意与不满意标准引人质量管理领域，并于1982年日本质量管理大会第12届年会上宣读了《魅力质量与必备质量》﹙Attractive Quality and Must-be Quality﹚的研究报告。 KANO模型定义了三个层次的顾客需求:基本型需求、期望型需求和兴奋型需求。这三种需求根据绩效指标分类就是基本因素、绩效因素和激励因素。 1. 卡诺模型简介 卡诺模型（KANO模型）是对用户需求分类和优先排序的有用工具，以分析用户需求对用户满意的影响为基础，体现了产品性能和用户满意之间的非线性关系。在卡诺模型中，将产品和服务的质量特性分为四种类型：⑴必备属性； ⑵期望属性；⑶魅力属性；⑷无差异属性。 ?魅力属性：用户意想不到的，如果不提供此需求，用户满意度不会降低，但当提供此需求，用户满意度会有很大提升； ?期望属性：当提供此需求，用户满意度会提升，当不提供此需求，用户满意度会降低； ?必备属性：当优化此需求，用户满意度不会提升，当不提供此需求，用户满意度会大幅降低； ?无差异因素：无论提供或不提供此需求，用户满意度都不会有改变，用户根本不在意； ?反向属性：用户根本都没有此需求，提供后用户满意度反而会下降 KANO问卷对每个质量特性都由正向和负向两个问题构成，分别测量用户在面对存在或不存在某项质量特性时的反应。

产品需求分析、满意度评价之KANO模型

产品需求分析、满意度评价之KANO模型KANO模型分析方法主要是通过标准化问卷进行调研,根据调研结果对各因素属性归类，解决产品属性的定位问题，以提高客户满意度的方法。 什么是KANO模型 KANO模型，是由东京理工大学教授狩野纪昭(Noriaki Kano)发明，用于分析用户对于各类需求的排名偏好情况，其在企业产品需求调研，市场研究中有着广泛的应用。 KANO模型的基本原理 KANO模型主要是通过标准化问卷进行调研，根据调研结果对各因素属性归类，KANO模型将功能/服务的态度属性分为六类，分别是魅力属性A、期望属性O、必备属性M、无差异属性I、反向属性R、可疑结果Q。

研究背景 当前有一项关于手机功能/服务的需求调研，共头脑风暴出10种功能/服务，分别是投影功能、左右手模式、超级快充、取消SIM卡、3D投影、照片搜索、自动美颜、防盗加锁、遥控器、暖手宝、望远镜、显微镜。 共收集有效数据为100份，现希望通过KANO模型分析出该10种功能/服务的态度情况，为企业产品研发提供建议。 问卷设计 设计KANO问卷时，针对每个功能需求，都需要设计正向和反向两个问题。 SPSSAU操作 ①选择SPSSAU【问卷研究】--【KANO模型】。

②将各功能指标的正反项放入对应分析框，同一题的正反两项放置的顺序需完全对应。 结果解读 1 KANO模型评价结果分类对照表

针对每个指标，KANO模型共分为正向题和负向题两个方向进行收集数据。并且在得到数据后，将指标映射到六个属性上。KANO模型评价结果分类对照表正是展现这样的对照表格。 注意：系统默认1分代表不喜欢，5分代表喜欢。如果你的数据不是这样设置的，可通过【数据处理】--【数据编码】进行修改。 2 KANO模型分析结果汇总 此表格为核心输出表格，即得出各功能/服务对应的属性占比、分类结果、Better和Worse值。

KANO模型

KANO模型 受行为科学家赫兹伯格的双因素理论的启发，东京理工大学教授狩野纪昭(Noriaki Kano)和他的同事Fumio Takahashi于1979年10月发表了《质量的保健因素和激励因素》(Motivator and Hygiene Factor in Quality)一文，第一次将满意与不满意标准引人质量管理领域，并于1982年日本质量管理大会第12届年会上宣读了《魅力质量与必备质量》﹙Attractive Quality and Must-be Quality﹚的研究报告。该论文于1984 年1月18日正式发表在日本质量管理学会(JSQC)的杂志《质量》总第l4期上，标志着狩野模式(Kano mode1)的确立和魅力质量理论的成熟。 kano模型 KANO模型简介 受行为科学家赫兹伯格的双因素理论的启发，东京理工大学教授狩野纪昭(Noriaki Kano)和他的同事Fumio Takahashi于1979年10月发表了《质量的保健因素和激励因素》(Motivator and Hygiene Factor in Quality)一文，第一次将满意与不满意标准引入质量管理领域，并于1982年日本质量管理大会第12届年会上宣读了《魅力质量与必备质量》﹙Attractive Quality and Must-be Quality﹚的研究报告。该论文于1984 年1月18日正式发表在日本质量管理学会(JSQC)的杂志《质量》总第l4期上，标志着狩野模式(Kano mode1)的确立和魅力质量理论的成熟。 KANO模型内容分析 KANO模型定义了三个层次的顾客需求:基本型需求、期望型需求和兴奋型需求。这三种需求根据绩效指标分类就是基本因素、绩效因素和激励因素。

kano模型

狩野模式(KANO Model) 目录 ? 1 KANO模型简介 ? 2 KANO模型内容分析 ? 3 KANO模型的实际操作意义 ? 4 KANO模型的优缺点分析 ? 5 KANO模型分析方法及应用实例[1] ? 6 相关条目 ?7 参考文献 KANO模型简介 受行为科学家赫兹伯格的双因素理论的启发，东京理工大学教授狩野纪昭(Noriaki Kano)和他的同事Fumio Takahashi于1979年10月发表了《质量的保健因素和激励因素》(Motivator and Hygiene Factor in Quality)一文，第一次将满意与不满意标准引人质量管理领域，并于1982年日本质量管理大会第12届年会上宣读了《魅力质量与必备质量》﹙Attractive Quality and Must-be Quality﹚的研究报告。该论文于1984 年1月18日正式发表在日本质量管理学会(JSQC)的杂志《质量》总第l4期上，标志着狩野模式(Kano mode1)的确立和魅力质量理论的成熟 Kano模型 KANO模型内容分析 KANO模型定义了三个层次的顾客需求:基本型需求、期望型需求和兴奋型需求。这三种需求根据绩效指标分类就是基本因素、绩效因素和激励因素。

基本型需求是顾客对企业提供的产品／服务因素的基本要求。这是顾客认为产品／服务“必须有”的属性或功能。当其特性不充足(不满足顾客需求)时，顾客很不满意；当其特性充足(满足顾客需求)时，顾客也可能不会因而表现出满意。对于基本型需求，即使超过了顾客的期望，但顾客充其量达到满意，不会对此表现出更多的好感。不过只要稍有一些疏忽，未达到顾客的期望，则顾客满意将一落千丈。对于顾客而言，这些需求是必须满足的，理所当然的。例如，夏天家庭使用空调，如果空调正常运行，顾客不会为此而对空调质量感到满意；反之，一旦空调出现问题，无法制冷，那么顾客对该品牌空调的满意水平则会明显下降，投诉、抱怨随之而来。 期望型需求是指顾客的满意状况与需求的满足程度成比例关系的需求。期望型需求没有基本型需求那样苛刻，其要求提供的产品／服务比较优秀，但并不是“必须”的产品属性或服务行为。企业提供的产品／服务水平超出顾客期望越多，顾客的满意状况越好，反之亦然。在市场调查中，顾客谈论的通常是期望型需求。质量投诉处理在我国的现状始终不令人满意，该服务也可以被视为期望型需求。如果企业对质量投诉处理得越圆满，那么顾客就越满意。 魅力型需求是指不会被顾客过分期望的需求。但魅力型需求一旦得到满足，顾客表现出的满意状况则也是非常高的。对于魅力型需求，随着满足顾客期望程度的增加，顾客满意也急剧上升；反之，即使在期望不满足时，顾客也不会因而表现出明显的不满意。这要求企业提供给顾客一些完全出乎意料的产品属性或服务行为，使顾客产生惊喜。顾客对一些产品／服务没有表达出明确的需求，当这些产品／服务提供给顾客时，顾客就会表现出非常满意，从而提高顾客的忠诚度。例如，一些著名品牌的企业能够定时进行产品的质量跟踪和回访，发布最新的产品信息和促销内容，并为顾客提供最便捷的购物方式。对此，即使另一些企业未提供这些服务，顾客也不会由此表现出不满意。 KANO模型的实际操作意义 在实际操作中，企业首先要全力以赴地满足顾客的基本型需求，保证顾客提出的问题得到认真的解决，重视顾客认为企业有义务做到的事情，尽量为顾客提供方便。以实现顾客最基本的需求满足。然后，企业应尽力去满足顾客的期望型需求，这是质量的竞争性因素。提供顾客喜爱的额外服务或产品功能，使其产品和服务优于竞争对手并有所不同，引导顾客加强对本企业的良好印象，使顾客达到满意。最后争取实现顾客的兴奋型需求，为企业建立最忠实地客户群。 以酒店行业为例，每种需求满意度如下： ★基本需求 清洁的床单正常工作的钥匙卡正确的帐单安全 ★“多多益善”的需求 早于承诺的时间将餐送到客人房间 优选房价 提供的服务符合品牌价值 ★“喜出望外”的需求 正确预计客人的需要,例如看到客人在咳嗽，员工能在客人要求之前，主动为客人送上一杯温开水 提供的服务与品牌价值相符。 KANO模型的优缺点分析

KANO模型

KANO模型定义了三个层次的顾客需求:基本型需求、期望型需求和兴奋型需求。这三种需求根据绩效指标分类就是基本因素、绩效因素和激励因素。 基本型需求是顾客认为产品“必须有”的属性或功能。当其特性不充 足（不满足顾客需求）时，顾客很不满意；当其特性充足（满足顾客需求）时，无所谓满意不满意，顾客充其量是满意。 期望型需求要求提供的产品或服务比较优秀，但并不是“必须”的产 品属性或服务行为有些期望型需求连顾客都不太清楚，但是是他们希望得 到的。在市场调查中，顾客谈论的通常是期望型需求，期望型需求在产品 中实现的越多，顾客就越满意;当没有满意这些需求时，顾客就不满意。 兴奋型需求要求提供给顾客一些完全出乎意料的产品属性或服务行为，使顾客产生惊喜。当其特性不充足时，并且是无关紧要的特性，则顾客无 所谓，当产品提供了这类需求中的服务时，顾客就会对产品非常满意，从 而提高顾客的忠诚度。 kano模型需求分析 1.基本品质（需求）。如果此类需求没有得到满足或表现欠佳，客户 的不满情绪会急剧增加，并且此类需求得到满足后，可以消除客户的不满，但并不能带来客户满意度的增加。产品的基本需求往往属于此类。对于这 类需求，企业的做法应该是注重不要在这方面失分。 kano模型

2.期望品质（需求）。此类需求得到满足或表现良好的话，客户满意度会显著增加，当此类需求得不到满足或表现不好的话，客户的不满也会显著增加。这是处于成长期的需求，客户、竞争对手和企业自身都关注的需求，也是体现竞争能力的需求。对于这类需求，企业的做法应该是注重提高这方面的质量，要力争超过竞争对手。 3.魅力品质（需求）。此类需求一经满足，即使表现并不完善，也能到来客户满意度的急剧提高，同时此类需求如果得不到满足，往往不会带来客户的不满。这类需求往往是代表顾客的潜在需求，企业的做法就是去寻找发掘这样的需求，领先对手。 KANO模型的实际操作意义 在实际操作中，企业首先耍全力以赴地满足顾客的基本型需求，保证顾客提出的问题得到认真的解决，重视顾客认为企业有义务做到的事情，尽量为顾客提供方便。以实现顾客最基本的需求满足。然后，企业应尽力去满足顾客的期望型需求，这是质量的竞争性因素。提供顾客喜爱的额外服务或产品功能，使其产品和服务优于竞争对手并有所不同，引导顾客加强对本企业的良好印象，使顾客达到满意。最后争取实现顾客的兴奋型需求，为企业建立最忠实地客户群。 kano模型

KANO模型简要分析

KANO模型简要分析 一、什么是KANO模型？ KANO模型分析方法是狩野纪昭基于KANO模型对顾客需求的细分原理，开发的一套结构型问卷和分析方法。 KANO模型分析方法主要是通过标准化问卷进行调研，根据调研结果对各因素属性归类，解决产品属性的定位问题，以提高客户满意度。 二、属性分类 在卡诺模型中，将产品功能/需求和服务的特性分为五种属性：必备属性、期望属性、魅力属性、无差异属性、反向属性。 必备属性：当优化此需求，用户满意度不会提升，当不提供此需求，用户满意度会大幅降低； 期望属性：当提供此需求，用户满意度会提升，当不提供此需求，用户满意度会降低； 魅力属性：用户意想不到的，如果不提供此需求，用户满意度不会降低，但当提供此需求，用户满意度会有很大提升； 无差异属性：无论提供或不提供此需求，用户满意度都不会有改变，用户根本不在意； 反向属性：用户根本都没有此需求，提供后用户满意度反而会下降；

根据KANO模型，将其属性分类与用户需求优先级进行对应，便于实际应用，主要定义了三种：基本型需求（必备属性）、期望型需求（期望属性）、兴奋型需求（魅力属性），这三种需求根据绩效指标分类就是基本因素、绩效因素和激励因素。 三、KANO模型实际操作流程 1.设计问卷调查表并实施有效的问卷调查 KANO问卷中每个属性特性都由正向和负向两个问题构成，分别测量用户在面对具备或不具备某项功能所做出的反应。问卷中的问题答案一般采用五级选

项，按照：喜欢、理应如此、无所谓、勉强接受、我不喜欢，进行评定，可以根据具体情况选择设置选项。 需要注意的点： ① KANO问卷中与每个功能点相关的题目都有正反两个问题，正反问题之间 的区别需注意强调，防止用户看错题意； ②功能的解释：简单描述该功能点，确保用户理解； ③选项说明：由于用户对“我很喜欢”“理应如此”“无所谓”“勉强接 受”“我很不喜欢”的理解不尽相同，因此需要在问卷填写前给出统一解释说明，让用户有一个相对一致的标准，方便填答。 我很喜欢：让你感到满意、开心、惊喜。 理应如此：你觉得是应该的、必备的功能/服务。 无所谓：你不会特别在意，但还可以接受。 勉强接受：你不喜欢，但是可以接受。 我很不喜欢：让你感到不满意。 2将调查结果的功能属性进行分类，建立原型 3确立功能影响程度究竟多大：Better-Worse系数-计算与使用 除了对于Kano属性归属的探讨，还可以通过对于功能属性归类的百分比，计算出Better-Worse系数，表示某功能可以增加满意或者消除不喜欢的影响程度。 计算公式如下：

Kano模型介绍及其应用

Kano模型介绍及其应用 满意度的二维模式 满意度是用户对产品感知的效果与期望值相比较后，用户形成的开心或失望的感觉。在日常满意度应用中，我们都认为满意度是一维的，即某个产品（页面），提供更多功能、服务时用户就会感到满意，相反，当功能、服务不充足时，用户会感到不满。因此我们可能会不断在产品（页面）中添加新功能，通过这种方式提升用户的满意度。但是事实上会发现，并不是所有新增或优化的功能，都能提升用户的满意度，甚至有一些还会损害用户体验。 满意度理论研究中发现，并非所有的因素对用户满意度产生的影响都是一维的，二维模式认为，当提供某些因素时，未必会获得用户的满意，有时可能会造成不满意，有时提供或不提供某些因素，用户认为根本无差异，这就是满意度的二维模式。 满意度的二维模式是从赫茨伯格（Herzberg）的双因素理论发展而来。赫茨伯格的理论认为，满意和不满意并非共存于单一的连续体中，而是截然分开的；该理论通过考察一群会计师和工程师的员工满意度与生产效率的关系，发现日常工作中员工的满意度分为两种，一种是激励因素，另一种称为保健因素。激励因素表示工作本身带来的成就、认可和责任；保健因素指公司政策和管理、技术监督、薪水、工作条件以及人际关系等。当具备激励因素时会增加员工的满意，但是当缺乏时不会不满意；而当具备保健因素时不会提高员工的满意，但是当缺乏时，则会造成不满。

Kano模型的二维属性模式 日本教授狩野纪昭(Noriaki Kano)在1984年首次提出二维模式，构建出kano 模型。将影响因素划分为五个类型，包括： 魅力因素：用户意想不到的，如果不提供此需求，用户满意度不会降低，但当提供此需求，用户满意度会有很大提升； 期望因素（一维因素）：当提供此需求，用户满意度会提升，当不提供此需求，用户满意度会降低； 必备因素：当优化此需求，用户满意度不会提升，当不提供此需求，用户满意度会大幅降低；无差异因素：无论提供或不提供此需求，用户满意度都不会有改变，用户根本不在意； 反向因素：用户根本都没有此需求，提供后用户满意度反而会下降； 从kano模型的因素分类可以发现，kano并不是直接用来测量用户满意度的方法，而是通过对用户的不同需求进行区分处理，帮助产品找出提高用户满意度的

Kano 模型与魅力质量理论综述

Kano模型和服务质量差距模型的比较研究 (魏丽坤兰州商学院 730020) [摘要] Kano模型和Gaps模型对顾客感知质量的研究各具特色但又异曲同工。比较两者的异同，整合其优势，并结合管理实践进行理论创新，对促进质量管理本土化，提高我国服务质量管理水平具有积极的意义。 [关键词]Kano模型；魅力质量；服务质量差距模型；SERVQUAL评价法；比较；创新； A Comparative Study of the Kano model and the Gaps Model of Service Quality LanZhou Commercial College WEI Li-kun Abstract：Kano model and Gaps model，each has particular but similar features and methods on the study of customer perceived Service quality .Comparing similarities and differences，integrating both advantages and making theory innovation link management practice. It has a positive significance to improve the service quality management level and nationalization in China. Key Words：Kano model；Attractive quality；Gaps model；SERVQUAL；Comparative Study 20世纪80年代，在质量管理领域产生了两种具有代表意义的管理方法，一种是以日本管理文化为背景的Kano模型和魅力质量理论，一种是以美国管理文化为背景的服务质量差距模型和SERVQUAL评价法，两种方法都极大地推动了顾客感知质量的研究和发展，并对后来的顾客满意、顾客忠诚管理产生了深远的影响。 1 Kano模型和魅力质量 1．1 发展历程 受行为科学赫兹伯格双因素理论的启发，东京理工大学教授狩野纪昭（Noriaki Kano）和他的同事Fumio Takahashi 于1979年10月发表了《质量的保健因素和激励因素》(Motivator and Hygiene Factor in Quality)一文，第一次将满意与没有满意标准引入到质量管理领域，并于1982年Nippon QC Gakka第12届年会上宣读了《魅力质量与必备质量》(Attractive Quality and Must-be Quality)的研究报告，该论文于1984 年1月18日正式发表在《日本质量控制协会》(Japanese Society Quality Control)杂志总14期上,标志着狩野模式（Kano model）的确立和魅力质量理论的成熟。1．2 主要内容 狩野教授引申双因素理论，对质量的认知也采用二维模式，即使用者主观感受与产品/服务客观表现，提出了著名的Kano 模型，如图1所示：

KANO模型在产品研发中的应用

卡诺模型（KANO） 在产品研发中的应用 产品部 谭军云 2016.07

KANO模型的价值 哪些功能该开发，哪些功能该舍弃？ 哪些问题应该优先解决，哪些可以推迟解决？ 哪些功能的用户体验提升对产品的用户满意度最有价值？

KANO模型介绍 ? 日本教授狩野纪昭(Noriaki Kano）受行为科学家赫兹伯格的双因素理论的启发，提出了满意度的二维模式-卡诺模型。? 二维模式认为，当提供某些因素时，未必会获得用户的满意，有时可能会造成不满意，有时提供或不提供某些因素， 用户认为根本无差异，这就是满意度的二维模式。 横坐标：代表用户需求的实现程度； 纵坐标：代表用户满意度； 红、绿、紫色线条代表用户满意度随着需求实现程度的变化情况；

需求分类 ? 对影响指标进行分类，对产品功能层次进行划分，帮助产品了解不同层次的用户需求，识别使用户满意的至关重要因素。 横坐标：代表用户需求的实现程度； 纵坐标：代表用户满意度； 红、绿、紫色线条代表用户满意度随着需求实现程度的变化情况；?兴奋型需求（A魅力因素）：是?户意想不到的产品或服务，使顾客产生惊喜。如果不提供此需求，用户满意度不会降低，但当提供此需求，用户满意度会有很大提升； ? 期望型需求（O一维因素）：是用户希望得到的但并不是“必须”的产品属性或服务。当提供此需求，用户满意度会提升，当不提供此需求，用户满意度会降低； ? 基本需求（M必备因素）：是用户认为产品“必须有”的属性或功能。当优化此需求，用户满意度不会提升，当不提供此需求，用户满意度会大幅降低； ?无关需求（I无差异因素）：无论提供或不提供此需求，用户满意度都不会有改变，用户根本不在意； ?反向需求（R反向因素）：用户根本都没有此需求，提供后用户满意度反而会下降； Kano模型从?用户?角度将需求划分为五种类型，从产品影响因素?角度也可划分为五类；

KANO模型详解word版本

K A N O模型详解

最早在腾讯的《在你身边为你设计》中看到这个模型，却一直没完全弄懂是怎么使用的，今天自己编造了一些数据，一步步做了一遍，总算理解了。 以下的引用部分引用自知乎。 1.卡诺模型简介-对用户满意度和需求进行分析的工具 卡诺模型（KANO模型）是对用户需求分类和优先排序的有用工具，以分析用户需求对用户满意的影响为基础，体现了产品性能和用户满意之间的非线性关系。在卡诺模型中，将产品和服务的质量特性分为四种类型：⑴必备属性；⑵期望属性；⑶魅力属性；⑷无差异属性。 KANO模型中的几种属性 魅力属性：用户意想不到的，如果不提供此需求，用户满意度不会降低，但当提供此需求，用户满意度会有很大提升； 期望属性：当提供此需求，用户满意度会提升，当不提供此需求，用户满意度会降低； 必备属性：当优化此需求，用户满意度不会提升，当不提供此需求，用户满意度会大幅降低； 无差异属性：无论提供或不提供此需求，用户满意度都不会有改变，用户根本不在意； 反向属性：用户根本都没有此需求，提供后用户满意度反而会下降

2.KANO模型的使用-问卷编制与数据处理 KANO问卷对每个质量特性都由正向和负向两个问题构成，分别测量用户在面对存在或不存在某项质量特性时的反应。需要注意： ① KANO问卷中与每个功能点相关的题目都有正反两个问题，正反问题之间的区别需注意强调，防止用户看错题意； ② 功能的解释：简单描述该功能点，确保用户理解； ③ 选项说明：由于用户对“我很喜欢”“理应如此”“无所谓”“勉强接受”“我很不喜欢”的理解不尽相同，因此需要在问卷填写前给出统一解释说明，让用户有一个相对一致的标准，方便填答。 我很喜欢：让你感到满意、开心、惊喜。 它理应如此：你觉得是应该的、必备的功能/服务。 无所谓：你不会特别在意，但还可以接受。 勉强接受：你不喜欢，但是可以接受。 我很不喜欢：让你感到不满意。 因此在编制问卷的时候，对每个项目都要有正反两道题来测，比如，“如果在微信中加入朋友圈功能，您怎样评价？”对应“如果在微信中去掉朋友圈功能，您怎样评价？”均提供五个选项：我很喜欢、它理应如此、无所谓、勉强接受、我很不喜欢 那么每个用户对于某一个项目的态度必然落入下图表中的某个格子。而对所有的用户来说，共有5*5即25种可能，统计每种可能下的用户人数占总人数的百分比，来填入下表。之后将下表中标A、O、M、I、R、Q的格子中百分比相加，即可得到五种属性对应的百分比。从需求的角度来说，先满足M百分比最高的去掉R百分比最高的，再满足O百分比最高的，最后满足A百分比最高的。