This paper highlights the simplicity and effectiveness of Computing with Words (CW) in the implementation of target selection. Direct marketing can be considered as one of the main areas of application for this methodology. In particular, fuzzy classification is applied in it with the purpose of choosing the best potential customers for a new product or service from a client database. One of the advantages of the proposed method is that it is consistent with relational databases. Our methodology makes it possible to form queries in natural language, such as “print the list of not very old married clients with more-or-less high income”, which is impossible using a standard query mechanism.
There is one fundamental advantage of humankind that
needs to be inculcated into the various information
systems. It is the remarkable human ability to perform a
wide range of mental tasks without any measurements and
any computations
The main purpose of using natural (linguistic) queries
instead of numbers is that it is much closer to the way that
humans express and use their knowledge. Perception-based
rational decisions in an environment of imprecision are
becoming highly actual
This paper is organized in six sections. First one is this
introduction. Next we emphasize the critical importance of
target selection in direct marketing. Furthermore, we
examine in details the how fuzzy approach was applied to
make the process of target selection more efficient.
Particularly, it discusses the concepts of linguistic
variables and hedges, fuzzification, explains the need for
Direct marketing (DM) is a form of advertising that enables
companies to communicate directly to the customer, with
various advertising techniques including email, mobile
messaging, promotional letters, etc. The crucial idea there
is to be able to deliver the marketing message to the clients
that are likely to be interested in the product, service, or
offer
An important data mining problem from the world of DM
is target selection
As mentioned above, in DM the selection of the target audience is a very important stage. Different DM techniques benefit from accurate target selection. Take, for example, a direct mail, used in the promotion of goods and services to organizations and individuals through electronic mail. Some DM methods using particular media, especially email have been criticized for poor target selection strategy. This poses a problem for marketers and consumers alike. On the one hand, advertisers do not wish to waste money on communicating with consumers not interested in their products. Also, they don’t want to lose potential customers. On the other hand, people usually try to avoid spam. However, they want to be aware of the new products/services that might be interesting for them.
As previously mentioned, in order to maximize its
benefits direct mail requires careful selection of recipients.
So, if the selection of recipients is too liberal, it will
increase unnecessary spending on DM, if it is too strict –
we’ll lose some potential customers. Virtually all
companies that work with a database of 100 or more
customers use email-mailing in their business
Nowadays, most of the data processed in information
systems has a precise nature. However, a query to the
database, formed by a person, often tends to have some
degree of fuzziness. For example, the result of a query in a
search engine is a lot of references to documents that are
ordered by the degree of relevance to the request. Another
simple example of natural query used in everyday life:
"Find a listing for housing that is not very expensive and is
close to downtown". Statements like "not very expensive,"
"close" are vague, imprecise, although rent price is
completely determined, and the distance from the center of
the apartment - up to a kilometer. The cause of all these
problems is that in real life, we operate and argue using
imprecise categories
For example, a company launches an advertising campaign among their clients about new services through direct mail. The Marketing Service has determined that the new service will be most interesting for middle-aged married men, with more-or-less high income. A crisp query, for example, might ask for all married males aged 40 to 55 with an income of more than 150,000 tenge. But with such a request we may weed out a lot of potential clients: a married man aged 39, with an income of 250,000 tenge does not fall into the query result, although he is a potential customer of the new service.
One of the main hindrances of modern computing is that a
concept cannot be well understood until it is expressed
quantitatively. This is where linguistic variables come in.
The main motivation to prefer linguistic variables rather
than numbers is that a linguistic description is usually less
specific than a numerical one
According to Zadeh , “By a linguistic variable we mean
a variable whose values are not numbers but words or
sentences in a natural or artificial language”
To clarify, in the example provided, Income is a linguistic variable, while low, average, high are linguistic values, represented in the form of fuzzy sets. The set of all linguistic values of a linguistic variable is called term set.
Although a linguistic value is less precise than a number
it is closer to human cognitive processes, and that can be
exploited successfully in solving problems involving
uncertain or ill-defined phenomena. So, in situations where
information is not precise (which are very common in our
real life), linguistic variables can be a powerful tool that
takes the human knowledge as model
Besides their primary meaning, linguistic values may involve connectives such as and, or, not and hedges such as very, quite extremely, more or less, completely, fairly, etc. about which we will talk extensively later.
It is highly important for any target selection model to
select the clients’ features that will play the role of
explanatory variables in the model
So, let’s suppose we have a table “Clients”, consisting of 7 rows: id (primary key), name, gender (‘Male’,‘Female’), age, status (‘Married’, ‘Not_married’), email, income.
By the way, in practice, a certain threshold of membership value is given in excess of which records are included in the result of a fuzzy query. Usually it is a number between 0 and 1 and can be represented to the user in the form of percentage. So, an expert can manoeuvre with it to make the query more or less strict. One of the situations in which a threshold can be very efficient is when expert receives a long list of clients as a response to a query. Then, he can decide to be stricter and make the threshold higher in order to be more confident in the buying power of the clients.
Usually, in real-world decision making processes there
are experts - decision makers who choose the appropriate
initial parameters to define the fuzzy variables
Many decision problems need to be solved under uncertain environments with blurred and imprecise information. The use of linguistic information in decision making involves processes of CW (discussed a bit later).
Fuzzy sets and logic play a major role in this project.
Fuzzy mathematics allows us to use the imprecision in a
positive way. It is very efficient in complex problems that
can’t be handled using standard mathematics, like
processing human elements – natural language, perception,
emotion, etc. The term “fuzzy” can be defined as “not
clear, blurred, or vague.” For example, the word “tall” is
fuzzy, since it is subjective term. For some people, man
with the height 190 cm (6.2 feet) is tall, whereas for others
170 cm (5.7 feet) is enough to call the person “tall”. As
Zadeh said, “Fuzzy logic is determined as a set of
mathematical principles for knowledge representation
based on degrees of membership rather than on the crisp
membership of classical binary logic”
A computationally efficient way to represent a fuzzy
number is to use the approach based on parameters of its
membership function. Linear trapezoidal or triangular
membership functions are good enough to catch the
ambiguity of the linguistic assessments
Now let’s try to formalize the fuzzy concept of the client’s age. This will be the name of the respective linguistic variable. We define it for the domain X = [0, 90], so, the universal set U = {0 ,1, 2, …..,89, 90}. The term set consists of 3 fuzzy sets – {"Young", "Middle-aged“, “Old“}.
The last thing left to do - to build certain membership functions belonging to each linguistic term – fuzzy set We define the membership functions for the young, middleaged, and old fuzzy sets with the following parameters [a,b,c] = [18,35,65]. In general form they look like:
Now we can, for example, calculate the degree of membership of a 30-year-old client in each of the fuzzy sets:
Another fuzzy variable in the system is client’s income. We define it for the domain X = [0, 1000 000], so, the universal set U = {0 ,1,…, 250 000, …,1 000 000}. The term set consists of 3 fuzzy sets – {"Low", "Average“, “High“}. The membership functions for income variable term set are totally similar to the ones discussed above. The parameters are the following [a,b,c] = [40 000, 100 000, 200 000]. Income fuzzy variable, so as Age, is partitioned by three fuzzy sets associated with linguistic labels. Each fuzzy set corresponds to perception agents – low, average, or high salary. As it can be seen from the graph, there are no sharp boundaries between low, average, and high.
It is highly important to remember that decision making is an inherent human ability which is not necessarily based on explicit assumptions or precise measurements. For example, typical decision making problem is to choose the best car to buy. Therefore, fuzzy sets theory can be applied to system to model the uncertainty of decision processes.
Knowledge representation through linguistic variables characterized by means of linguistic modifiers – hedges makes the query more natural, so their main advantage is the ability to be expresses in natural language. Hedges can change the statement in various ways – intensify, weaken, complement. Their meaning implicitly involves fuzziness, so their primary job is to make things fuzzier or less fuzzy.
Let’s consider the most common ways of generating new
fuzzy sets based on the initial fuzzy set using various
hedges. This is useful for constructing various semantic
structures -composite words - from atomic words (i.e.
young) that reinforce or weaken the statements
Again, the main motivation is to strengthen or weaken
the statement
For example, let’s suppose that the meaning of X
(middle-aged) is defined by some membership function. If
we want to strengthen the statement, we use very
intensifier
Furthermore, the modifier that can weaken the statement
- more-or-less X (i.e. more-or-less middle-aged) would be
given as a square root of the initial function
Finally, not X (i.e. not young) which is a complement fuzzy set, can be expressed by subtracting the membership function of X (middle-aged) from 1: μ F NOT ( X ) = 1 − μ F ( X )
Let’s enjoy calculating the membership of 30-year-old client to each of the fuzzy sets: middle-aged, not middleaged, very middle-aged, and more-or-less middle-aged.
As we have seen, hedges intrinsically convey the imprecision in themselves. The main flexibility they provide is that they can make a fuzzy natural query even more natural.
Computing with words (CW), originally developed by
Zadeh, provides a much more expressive language for
knowledge representation. In it, words are used in place of
numbers for computing and reasoning, with the fuzzy set
playing the role of a fuzzy constraint on a variable
A basic premise in CW is that the meaning of a
proposition, p, may be expressed as a generalized
constraint in which the constrained variable and the
constraining relation are, in general, implicit in p
CW methodology is changed a little bit to correspond to the proposed system. In particular, the initial data set (IDS) is a database with clients’ information. From the IDS we desire to find the subset of clients from the database in response to a query expressed in a natural language. That is our result - terminal data set (TDS). So, our goal is to derive TDS from IDS. In our model, we process the natural query step by step, by constraining the values of variables, this process will be considered in details later.
Our aim is to make explicit the implicit fuzzy constraints which are resident in a query. So, how can we make explicit the fuzzy constraints that are given in natural language and so as are implicit?
In linguistic features of variables, words play the role of
the values of variables and serve as fuzzy constraints at the
same time
In outline, a query q in a natural language can be considered as a network of fuzzy constraints. After processing procedure we get a number of overall fuzzy constraints, which can be represented in the form X is R, Y is S…, where X is a constrained criterion variable (i.e. age) which is not explicit in q, and R is a constraint on that criterion. So the explicitation process can be defined as: q → X is R, Y is S…, etc.
As a simple illustration, let’s consider the simple query: not very young males with more-or-less high income. As we can observe, some of the variables in a query are crisp, while some have fuzzy constraints. Let’s assume that the user chose the threshold value μ Total as a sufficient level of precision. So, we obtain:
Notice that not very young and very not young are different things. Therefore, the order is important. Another main point is that μTotal is a membership value reflecting the degree of membership to not very young and more-orless high, not to young and high. We need to pay special attention to it. To obtain the answer we need the membership value that corresponds to young and high, of course. That is why, the process is reversed: before we presented the formulas to shift to very young from young. Now, instead, we want to define young using very young. So, if we squared the threshold for young to get the threshold for very young, now we apply the inverse operation – square root. Furthermore, we get:
∩ MALE[Gender; ; μ Total = 1] = YOUNG[Age; ;
Let’s consider the translation rules that can be applied
singly or in combination
Constraint Modification Rules. X is mA → X is f (A) where m is a modifier – hedge or negation (very, not, more-or-less), and f (A) defines the way m modifies A.
It should be stressed that the rule represented is a
convention and shouldn’t be considered as the exact
reflection of how very, not or more-or-less function in a
natural language
Constraint Propagation Rules. Constraint propagation plays crucial role in CW. It is great that all the stuff with numbers takes plays outside of a user’s vision.
The rule governing fuzzy constraint propagation: If A
and B are fuzzy relations, then disjunction – or (union) and
conjunction – and (intersection) are defined, respectively,
as max and min
Users can express the intersection in 3 ways distinguished by the connective type – and, but, or no connective at all. As it was previously stated, for this operation, we take the minimum of two memberships to get the resultant membership value:
μA( x) ∩ B( x) = min[μA( x), μB( x)]
Union is represented solely by or connective. The resultant membership value is equal to the maximum of two values provided:
μA(x) ∪ B( x) = max[μA( x), μB( x)]
The threshold used in the system serves as the α-cut (Alpha cut), which is a crisp set that includes all the members of the given fuzzy subset f whose values are not less than α for 0< α ≤ 1:
fα = { x : μ f ( x ) ≥ α }
We also know how to connect α-cuts and set operations (let A and B be fuzzy sets):
( A ∪ B)α = Aα ∪ Bα , ( A ∩ B)α = Aα ∩ Bα
So, using the formulas provided above, in order to find the result of a query with a certain threshold – α, containing or or and operations, we first find the α-cuts and then take the crisp or / and operation.
In dealing with real-world problems there is much to be
gained by exploiting the tolerance for imprecision,
uncertainty and partial truth. This is the primary
motivation for the methodology of CW
More and more employees are depending on information from databases to fulfill everyday tasks. That is why nowadays it is becoming increasingly important to access information in a more human-oriented way – using natural language.
We presented a fuzzy querying model capable of handling various types of requests in a natural language form. The interface developed allows experts to express questions in natural language and to obtain answers in a readable style, while modifying neither the structure of the database nor the database management system (DBMS) query language.
The main advantage of our model is that the query to the system is done in a natural language. Besides, existing clients databases do not have to be modified and developers do not have to learn a new query language. Basically, we just have a fuzzy interface that is used as a top layer, on an existing relational database, so, no modifications on its DBMS were done.
The main problem in developing human-oriented query
interfaces was how to allow users to query databases in a
natural way. The motivation for that is that usually users
don’t wish to define the clear bounds of acceptance or
rejection for a condition, that is, they want to be allowed
some imprecision in the query
The past and new conceptual structure of the model is schematized and illustrated below, in figure 10.
Let’s look at examples of natural language queries given with the purpose of demonstrating the capabilities of this human-oriented interface.
The application interface is very friendly. If some user experiences problems forming a natural query, he can use another menu provided. In it the criteria are listed on the screen, and user can just pick and choose which ones he wants. Furthermore, the parameters he chooses appear and he needs to choose needed values (“very young”, “not old”, etc.) on the respective pull-down menu. Moreover, in order to adapt this model to other databases we won’t need to change the logic, because it is context-independent. We will just need to change the list of fuzzy variables.
Example query 1. not old married males with very high income. [Threshold value: 0.5]
Here we have two crisp criteria - status is married, gender is male. Furthermore, there are two fuzzy criteria – age is not old and income is very high. So, we have: OLD[Age; not; μ Total =0.5] ∩ HIGH[Income; very; μ Total =0.5] ∩ MALE[Gender; ; μ Total = 1] ∩ MARRIED[Status; ; μ Total = 1] = OLD[Age;; μ Total =0.5] ∩ HIGH[Income; very; μ Total ≈ 0.7 ] ∩ MALE[Gender; ; μ Total = 1] ∩ MARRIED[Status;; μ Total = 1]
Next our system finds the values of age and income that correspond to the thresholds obtained. For the age, the constraining relation will be “ ≤ 50 ”, for the income – “ ≥ 170 710 tg.”.
Now, having a look at our sample table, we can find 2 clients, fully satisfying the query. The system gives us the same result: id name genderage status email income 5 Ernar M. Male 32 Married era@gmail.... 890 009 8 Karl L. Male 50 Married karl@hotm... 200 300
Example query 2. middle-aged but not more-or-less old clients. [Threshold value: 0.5]
Here we need to make the conjunction of two constraints on one fuzzy variable – age: MIDDLE-AGED[Age; ; μ Total =0.5] ∩ OLD[Age; not, more-or-less; μ Total =0.5] = MIDDLE-AGED[Age; ; μ Total =0.5] ∩ OLD[Age; ; μ Total =0.25]
We obtain the following result (note, that if we queried just for middle-aged, then 50-yeared client Karl L. would be included to the result set): id name gender age status email income 4 Iliyas T. Male 28 Not_mar.. iliyas@gma... 305 000 5 Ernar M. Male 32 Married era@gmail.... 890 009 6 Kamin... Female 40 Married kaminari@... 55 000 11Madina.. Female 34 Not_mar… madina_@... 30 000
Example query 3. not very young married clients with average or more-or-less high salary. [Threshold value: 0.7] We obtain the following: YOUNG[Age; not, very; μ Total =0.7] ∩ MARRIED[STATUS; ; μ Total =1] ∩ (AVERAGE[Income; ; μ Total =0.7] ∪ HIGH[Income; more-or-less ; μ Total =0.7]) =
The result set is the following: id name gender age status email income 5 Ernar M. Male 32 Married era@gmail.... 890 009 8 Karl L. Male 50 Married karl@hotm... 200 300 9 Amina L. Female 74 Married amina@ya... 120 000 13 Alfi A. Male 67 Married alfi@gmail... 88 000 Last thing to note, the hedges can be applied infinitely in any order! In order to demonstrate that in practice, consider the following example.
Example query 4. very very very old or very very very young. [Threshold value: 0.5] OLD[Age; very, very, very ; μ Total =0.5] ∪ YOUNG[Age; very, very, very; μ Total =0.5] = OLD[Age; ; μ Total ≈ 0.92] ∪ YOUNG[Age; ; μ Total ≈ 0.92]
The targeted clients are: id name gender age 9 Amina L. Female 74 10 Alan D. Male 18 13 Alfi A. Male 67 status email income Married amina@ya... 120 000 Not_mar alan@gmai... 35 000
Married alfi@gmail... 88 000
For sure, such type of human oriented interfaces can be very useful for all companies that face the problem of efficient target selection of clients.
There is one thing that disorients me in our Direct Marketing System. Using Zadeh's definition of the very intensifier it follows that the curve for very young, must hit the values 0 and 1 at exactly the same places as the curve for young. It is counterintuitive in this particular application (as well as others), since it can be absolutely possible that someone is young without it being absolutely true that he is very young. This contradiction, no doubts, gets even worse with very very young, very very very young, etc. According to Zadeh's, they all hit the values 1 and 0 at the same place as young.
A different model for the hedges may likely be necessary for a future improvement, that narrows down the range of ‘very young’ whose membership values are 1 when comparing with that of ‘young’ for example.
The main goal of this research was to demonstrate the effectiveness of Computing with Words approach in natural query processing. In a nutshell, it allows us to form queries in natural language, which is impossible using a standard query mechanism, thus simplifying the life for an expert.
In certain areas, like Direct Marketing, target selection of information from databases has very blurred conditions. Fuzzy queries can be very efficient there. Similarly, fuzzy queries can be used in the variety of other fields. Namely, in selecting tourist services, real estate, etc.
To conclude, the use of natural language in decision problems is highly beneficial when the values cannot be expressed by means of numerical values. That happens quite often, since in natural language, truth is a matter of degree, not an absolute.
There are future improvements. However, those are mostly some minor technicality such as the matter of linguistic hedges being counterintuitive and some auxiliary, cosmetic functionality such as a parser and GUI when considering some system development.