Download pptx - Tender Miner

Transcript
Page 1: Tender Miner

Tender Analysis Module

Paras Deshpande

mmpd2
Page 2: Tender Miner

Tenders

The tenders have various attributes like country, document type, CPV codes.

These attributes can have different values depending upon the particular tender in question.

Also the tender can have same attribute with different values e.g. CPV Codes.

Page 3: Tender Miner

Methodology: Analysis of Tenders

The attributes of the tenders have been formulated as Categories. These are therefore general representation of the attributes.

The various specific values that the attributes takes are captured into criteria. The tender thus has a dual level one with a broader scope while the other encapsulates the specific information.

Page 4: Tender Miner

Methodology: Analysis of Tenders cont

The more categories a tender has the more attributes it will be evaluated upon.

The more criteria's a category has the category will become more comprehensive and general.

The ranking given to a criteria is normalised over other criteria within the category.

Page 5: Tender Miner

Advantages of Weighted Criteria Algorithm

The Algorithm offers various advantages enumerated below:

Firstly, the Algorithm is simple and intuitive yet very effective in case of Tender Analysis.

The Algorithm is scalable and can be robustly implemented for very large number of Tenders.

Page 6: Tender Miner

AHP vis-à-vis Weighted Criteria

AHP is effective when decisions involve many intangibles to be measured along side tangibles whose measurements must also be evaluated by the decision maker

Therefore when the measurement of the criteria’s assumes subjectivity then AHP is the method of choice.

In case of Tender Analysis the measurement of the criteria’s is tangible and hence the Weighted Criteria method preferred apart from the scalability issues.

Page 7: Tender Miner

Mathematical formulation

Page 8: Tender Miner

Notations

{1,..., }

{1,..., }

{1,..., . }

{1,...

i

i I where I T set of candidate tenders

j J where J P set of all the attributes of tender i

a C whereC no of Categories set of all the categories of profile

b c where c

, }aC set of all the criterias of a category

Page 9: Tender Miner

Weightages

a

ba

W weightage of Category a for the profile

w weightage of criteria b for the Category a

, ' /

, ' /

a a a

ba ba ba

Normalized weightage of Category a for the profile W W W

Normalized weightage of Criteria b for the Category a w w w

• We then Normalise the weightages to the scale of -1 to +1

Page 10: Tender Miner

Calculation of Tender Score

1

1

1,

0

, ' .

, ' .

bai

c

ai ba baib

C

a aia

if the criteria b is satisfied for tender iCriteria Score cs

otherwise

Category Score CS w cs

Tender Score TSi W CS

Page 11: Tender Miner

Preferencei

Page 12: Tender Miner

Need for iPreference

iPreference stands for inferred preferences. iPreference is the logical extension of the

Tender Miner because of its unique capability to infer, map and store user preferences.

Page 13: Tender Miner

Features of iPreference

User can now store previously won tenders as favorites which will subsequently be used to infer their preferences and hence rank tenders according to inferred preferences.

This will greatly improve the user experience by saving him from cumbersome task of filling the forms and also address changing preferences.

Page 14: Tender Miner

Methodology

The User will be required to fill in the categories he wishes to have the tenders judged upon along with their weightages.

He then has a option of either judging some tenders provided by the service OR can add a list of previously won tenders to his favorites.

Now the iPreference uses these tenders to infer

user preference and adds his inferred profile to the knowledge Base.

Page 15: Tender Miner

Methodology cont

The logic used for inferring is described next: Ranking Tenders requires categories and their

weightages, criteria and their weightages. In this case instead of criteria we have tenders

with their weightages. Therefore a simple correlation logic would

suggest: More favorable a tender is More favorable its criteria's are expected to be.

Page 16: Tender Miner

Advantages

This is the first* time a reverse approach is being taken for decision making.

Its reverse in a sense it presents the reference alternatives and then ranks candidate alternatives on the basis of knowledge gained rather than taking down the preferences first.

* To best of my knowledge

Page 17: Tender Miner

Finer Aspects

There are some issues which need further attention. For e.g. if a favorable tender might have some favorable and some unfavorable criteria's, and presently iPreference cannot differentiate between them.

It may be agued though as more tenders and fairly representative tenders are added to the list the unfavorable criteria’s will slowly get neglected as happens with heuristics.

Eventually the better criteria will prevail !!!

Page 18: Tender Miner

Finer Aspects cont

This method is similar to Neural Networks in many cases and has similar advantages and limitations.

This method requires less effort on learning routines.

Under learnt profile might result in substandard results and a minimum of 10-15 tenders should be included in favorites.

Page 19: Tender Miner

Mathematical Formulations

Page 20: Tender Miner

Notations

{1,..., }

{1,..., }

{1,..., } ' '

{1,..., . }

R

i

r R where R T set of reference tenders

i I where I T set of candidate tenders

j J where J P set of all the attributes of tender i

a A where A no of Categories set of

' ' infe

' '

red

a a

let b be the index of criterias that wi

all the categories of profile

b B where B is set

ll be

of all the criterias of category a

Page 21: Tender Miner

Weightages

' '

' ' expert

infered

a

ba

W weightage of Category a for the profile

Wr weightage of Tender r given by the

w weightage of criteria b for the Category a

' ' '

' ' ' expertaW Normalized weightage of Category a for the profile

W r Normalized weightage of Tender r given by the

• Normalized weightages :

Page 22: Tender Miner

Algorithm

1 ' ' ' ' ' '

0

.

' infered ' ' ' '

R

rba

T

ba rbar

ba

if the the tender r has citeria b of category ax

otherwise

w x Wr

w Normalised weightage of criteria b for the Category a

Page 23: Tender Miner

Ending Remarks

The Algorithm shows that with different types of tenders being added to the favorites the favorable criteria will occur more frequently and thus grow stronger in weightages.

There can be improvements on finding ways to promote favorable criteria's and eliminating the unfavorable ones.

The main contribution thus will be proposing a intuitive and simple inference mechanism.

Page 24: Tender Miner

References

http://people.revoledu.com/kardi/tutorial/AHP/Multi-Criteria-Decision-Making.htm

Saaty, T.L. (2008) ‘Decision making with the analytic hierarchy process’, Int. J. Services Sciences, Vol. 1, No. 1, pp.83–98.

Page 25: Tender Miner

End of Slides