FINDING NON-OBVIOUS PROFILES BY USING

ANT-ALGORITHMS

Thomas Ambrosi

Information Technology, Nortel GmbH, D-60549 Frankfurt/Main, Germany

Sascha Kaufmann

FSTC/ILIAS/MINE; Universit

e du Luxembourg, 6, rue R. Coudenhove-Kalergi, L-1359 Luxembourg, Luxembourg

Keywords:

User interest proﬁles, Recommendations, Ant-algorithms.

Abstract:

Visitors on a website are usually on their own when they are moving around. First time visitors are especially

guessing where to ﬁnd the information they are looking for. In this paper we will show a way to combine

the concepts of non-obvious proﬁles and ant-algorithms to come up with a set of paths, which tries to cover

the user’s interests in a proper way. These paths can be used to give recommendations to visitors. While the

proﬁles help to get a better understanding of the users’ interests, the concept of ant-algorithms is employed to

determine recently and frequently used paths to lead the user to the desired information. We explain the basic

idea of our approach, the current state of the prototype realized and some ﬁrst results.

1 INTRODUCTION

There are many factors that inﬂuence a user’s expe-

rience when visiting a website (e.g. GUI, content,

structure). In this work we concentrate on a web-site’s

structure. We focus on content-directed users, which

are looking for interesting information. For this rea-

son, we are assuming, that visitors want to minimise

the number of pages to be visited during their atten-

dance which revise stress and gives a good surﬁng ex-

perience. We want to achieve this by performing a

personalised optimisation.

The work is partitioned into the following sec-

tions: In section 2 we describe the related work. In

section 3 we present our new approach. Section 4

contains ﬁrst results, while section 5 gives a summary

and an outlook to further work.

2 RELATED WORK

Our approach combines two algorithms: the NOP-

Algorithm and the concept of ant-algorithms.

The concept of “Non-Obvious Proﬁles (NOP)”

was initially designed to measure the level of inter-

ests in certain topics of web portal users over a period

of time (Mushtaq et al., 2004). In an initial process,

every page of a website is assigned one or more com-

binations of weighted topics, that indicate how well

a topic represents the page’s content. This is used

later to calculate an interest proﬁle each time a per-

son visits the website. The algorithm is doing this by

only taking into account how long the user stays at

each page during the session. As an adjusting mech-

anism, the user is asked from time to time about his

current interests. The results are used to calibrate the

formerly calculated proﬁles. The algorithm was later

extended by adding the concepts of areas and actions

to improve the proﬁle’s quality (Hoebel et al., 2006).

The ant-algorithm is based on experiments done

on ants (e.g. (Goss et al., 1989), (Jeanson et al.,

2003)), which observed them in their natural environ-

ment and how they communicated via pheromones.

Pheromones are scents that are secreted by ants. They

create a kind of path of scent that other ants can fol-

low. Over time, depending on the kind of surface,

the scent vanishes if it is not renewed. Pheromones

are employed for navigation during different actions

(e.g. searching for food). In computer-science ar-

tiﬁcial ant-algorithms are mainly used for optimisa-

tion problems, as the ”Travelling Salesman Problem

(TSP)” (Bonabeau et al., 1999).

681

Ambrosi T. and Kaufmann S.

FINDING NON-OBVIOUS PROFILES BY USING ANT-ALGORITHMS.

DOI: 10.5220/0001842806740679

In Proceedings of the Fifth International Conference on Web Information Systems and Technologies (WEBIST 2009), page

ISBN: 978-989-8111-81-4

3 OUR NEW APPROACH

As presented, our approach combines the NOP’s char-

acteristic of interesting proﬁles with the ant approach

of handling paths. We use the characteristic of a pos-

itive reinforcement of short ways, which is presented

by the natural trace of pheromones.

Like in the TSP, we model the covered distance

as a directed graph. Adding the according pheromone

assignment to the appropriate accumulations will then

become visible.

In our work, we refer to the “Adaptive Colony

System (ACS)-TSP algorithm” (Bonabeau et al.,

1999). The ant covers a distance during its search for

food and its return journey. It marks its path with a

trace of one or more different pheromones. Every ant

is able to orient itself using these traces. The trace’s

strength will be important when the ant’s path splits.

In most cases, the ant will follow the stronger trace,

but its decision can be inﬂuenced by other parameters,

too.

In our model a visitor adopts the ants role and

has the ability to dispense and recognise a special

pheromone for each topic (i.e. interests of the visi-

tor). The released amount of a pheromone is related

to the time a visitor is on a page and the topics (along

with the weights), which are related to this page. It is

also decreasing with any additional visited page. The

ant is able to recognise existing traces of pheromones

and store them as a non-obvious proﬁle. These traces

can be given by the website administrator or by other

visitors.

Applications of our combined approach could be

to indicate visitors’ interests in a faster way and sug-

gest web pages that might cover their interest proﬁles

with a high probability. We use the pheromone trace

to analyse the website for typical patterns of paths

taken by visitors, based on their interests.

3.1 Distances

As shown in (Traniello, 1989), ants do not use ar-

bitrary ways to get their food. Usually they try to

minimise the distance. We consider this by intro-

ducing a factor PherDist, which is steadily decreas-

ing with each additional page. As already mentioned

we assume, that after a large number of requests, the

user is not interested anymore in the content (content-

directed visitor). We describe this scenario as follows:

• The ﬁrst pages are very important

• The importance is constantly decreasing

We have experimented with two different distance

functions to inﬂuence a user proﬁle. We are using

equation (1) to express an exponential decreasing,

where l is the length of the path, i.e. number of pages,

and f is a factor controlling the aging rate.

PherDist(l) = e

f ∗l

where f ≤ 0 (1)

Alternatively, we are suggesting a function with

a linear aging rate that prunes the inﬂuence of later

pages at one point:

PherDist(l) =



−

+ 1 where 0 < l ≤ f

0 otherwise

(2)

In practice, these functions depend on the purpose

of the website. Other functions can be considered as

long as they satisfy the condition (3):

0 ≤ PherDist(l) ≤ 1, l = 1, .. . , l

max

(3)

3.2 Aging of Pheromones

While the information of knowing where to ﬁnd the

food is very important, it is as important to know

when a path was last used. The assumption here is to

recommend a path that was taken by a visitor with the

same interests in the recent history. Here we use the

fact that the strength of the pheromones are steadily

decreasing over time (Jeanson et al., 2003). After

each iteration the pheromones’ strengths are decreas-

ing by a factor PherAge. Similar to (Bonabeau et al.,

1999) we use an exponential decreasing function (4)

for our experiments, where w

is the visitor’s current

value of interest for topic k, f is a factor and t is the

number of iteration between the last computation of

the visitor’s proﬁle and the current one.

PherAge(t) = w

∗ e

f ∗t

where f ≤ 0 (4)

As an alternative we used function (5) in our ex-

periments which has a different behaviour.

PherAge(t) = f ∗ w

where f ≤ 0 (5)

The deﬁnition of an iteration is very important and

therefore the understanding of ’recent past’. In our

work we assume that all actions that occur in an iter-

ation take place simultaneously. In practice it highly

depends on the amount of incoming users in a time

span. An appropriate value for PherAge(t) should sat-

isfy condition (6):

0 ≤ PherAge(t) ≤ 1, t = 1, . . . ,t

max

(6)

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

682

3.3 Calculating the Extended NOP

We consider a website as a directed graph G = (N, E),

where N is the set of all pages. The directed edges E

are all v

i j

where j is directly reachable from i. We

assign all topics and their weights of page P

to the

edges v

i j

and the pheromones are represented by the

topics.

A user’s visit is represented as a sequence of

edges v

, ..., v

. We are then able to calculate the

k-pheromone’s contribution τ

, to the non-obvious

proﬁle’s value w

for topic k by using (7), where

duration(v

) is the time the user stayed on Page P

Note that using the duration of the whole session nor-

malises the pheromones’ sums.

∆τ

∑

i=1

duration(V

) · PherDist(i) · v

(τ

)

∑

j=1

duration(V

)

(7)

In the next step the pheromones sum is added

to the existing non-obvious user proﬁle. As the

pheromones’ values decrease over time, we calculate

the non-obvious proﬁle as follows (8):

= w

· PherAge(Date(w

))+

SessionCount

∑

i=1

PherAge(SessionDate(i)) ·

∆τ

(Session

)

(8)

Here, w

is the value that represents the users in-

terest in topic k. Date and SessionDate represent the

time and the number of iterations respecting since the

last computation of the non-obvious proﬁle.

4 RESULTS

We experimented with data from three existing web-

sites, however, in this paper we present only the web-

site with the best results. The reason for this is that

we had to assign topics manually by reading the con-

tent. This website was relatively ’small’ (approx.

400 pages) and static. It allowed us to do the topic-

weight assignment in an accurate way, while the other

two websites were either very large and/or already

changed by the time we did the assignment.

4.1 Testing Environment

The websites were not designed to support the con-

cept of non obvious proﬁles. A feedback mechanism,

to calibrate the visitor’s interests, was also not given.

We used the anonymized log ﬁle of a commercial

website with removed known robots’ entries (Cooley

et al., 1999). We then deﬁned ten topics, which de-

scribed in our understanding the website in a good

way and manually assigned the relevant topics to ev-

ery page. Note that this process is very subjective and

is highly inﬂuenced by the person who is doing the

assignments. Table (1) tabulates our test data parame-

ters, whilst Figure (1) gives an overview of the distri-

bution of the number of pages per session.

Table 1: Testing environment parameters.

Parameter Value

Observed Period 09/27/2005 -

12/27/2005

# Entries 57 785

# Sessions 6 227

# Sessions ≥ 3 requests 3 887

∅ Session Length (≥ 3) 14.115

Filters Elimination of

known robots

# Topics 10

Figure 1: Distribution no. of pages vs. no. of session.

We determined the exposure time ∆t for each page

by comparing the times when the web pages were re-

quested (log ﬁle). We calculated ∆t using (9), where

is the point in time when page 1 was sent, and t

the point in time when page 0 was sent, with t

≥ t

∆t = t

− t

(9)

To identify a user during a session we used the ip-

address or, in case of existence, a session id. Further-

more we used a session time-out of 24 minutes for all

pages, because it was already deﬁned on one website.

Given this data, we came up with a set of click

streams. A click stream represents a directed graph

= (N, E), where G

is a set of visited pages. The

directed edges are all v

with n being the page fol-

lowing m. Finally, the exposure time is added as an

attribute to the outgoing edges.

FINDING NON-OBVIOUS PROFILES BY USING ANT-ALGORITHMS

683

4.2 Test Results

In the following we show the results for the ob-

tained proﬁles by applying the algorithm. To com-

pare the different results in a better way, we com-

puted the maximum and the average values for each

pheromone/topic based on the used iteration time

span. The average pheromone values were computed

by all proﬁles with more than one session. The rea-

son for this was not to inﬂuence these values by too

many ’trivial’ NOPs and to get a better understanding

of multiple visits.

For getting a feeling of the level of inﬂuence of

the PherAge-function we experimented with two dif-

ferent functions. We used the values f = 0.8, f =

0.9, f = 0.99 in conjunction with equation (5) while

the values f = −0.002, f = −0.00025 indicates that

we used equation (4).

4.2.1 Results Iteration Size 24 Hours

Table 2 shows the average values in the results for

an iteration time span of 24 hours. This values seem

good, but a closer look at the maximum values for the

topics in Table 3 shows that the values can rise very

high in comparison with the original NOP-approach.

This happens if a visitor comes back within 24 hours

(multiple sessions) there will be a direct addition of

the iteration’s proﬁle values.

There are two disadvantages with this solution: it

makes it difﬁcult to compare the values and we can

easily become dependent on a few users. Promising

modiﬁcation of the algorithm can be to decrease the

interval’s time span or to satiate the pheromones’ val-

ues if there are multiple sessions of the same visitor

during an iteration.

Table 2: NOPs average values (24 h).

Avg. f=0.8 f=0.9 f=0.99 f=-0.002 f=-2.5*10

-4

P 01 0.25 0.26 0.30 0.28 0.30

P 02 0.07 0.08 0.09 0.08 0.09

P 03 0.11 0.11 0.13 0.12 0.13

P 04 0.09 0.10 0.12 0.11 0.12

P 05 0.07 0.07 0.09 0.08 0.09

P 06 0.17 0.17 0.20 0.18 0.20

P 07 0.17 0.18 0.20 0.18 0.20

P 08 0.56 0.58 0.64 0.60 0.64

P 09 0.69 0.71 0.79 0.74 0.80

P 10 0.52 0.54 0.62 0.58 0.63

4.2.2 Results Iteration Size 6 Hours

In the next step we reduced the iterations’ time span

to 6 hours. There were no real improvement, the val-

Table 3: NOPs maximum values (24 h).

Max. f=0.8 f=0.9 f=0.99 f=-0.002 f=-2.5*10

-4

P 01 1.57 1.57 1.77 1.71 1.81

P 02 0.59 0.59 0.59 0.59 0.59

P 03 1.12 1.12 1.12 1.12 1.12

P 04 0.88 0.88 0.88 0.88 0.88

P 05 0.98 0.98 0.98 0.98 0.98

P 06 1.35 1.35 1.35 1.35 1.35

P 07 1.32 1.32 1.32 1.32 1.32

P 08 3.19 3.34 3.81 3.48 3.90

P 09 2.65 2.90 4.77 3.28 4.85

P 10 2.72 2.72 2.72 2.72 2.72

ues were nearly the same as in the 24 hours iteration

interval (Tables 4 and 5).

This indicates a constant interest in a topic and

that visitors with multiple sessions within 6 hours, be-

ing constantly interested in the same kind of topics.

Table 4: NOPs average values (6 h).

Avg. f=0.8 f=0.9 f=0.99 f=-0.002 f=-2.5*10

-4

P 01 0.24 0.25 0.28 0.26 0.27

P 02 0.07 0.07 0.08 0.07 0.08

P 03 0.10 0.11 0.12 0.11 0.12

P 04 0.08 0.09 0.11 0.09 0.10

P 05 0.06 0.07 0.08 0.07 0.08

P 06 0.16 0.16 0.18 0.17 0.18

P 07 0.16 0.17 0.18 0.17 0.18

P 08 0.54 0.56 0.60 0.57 0.59

P 09 0.65 0.67 0.74 0.70 0.72

P 10 0.49 0.51 0.58 0.54 0.57

Table 5: NOPs maximum values (6 h).

Max. f=0.8 f=0.9 f=0.99 f=-0.002 f=-2.5*10

-4

P 01 1.57 1.57 1.64 1.57 1.69

P 02 0.59 0.59 0.59 0.59 0.59

P 03 1.12 1.12 1.12 1.12 1.12

P 04 0.88 0.88 0.88 0.88 0.88

P 05 0.98 0.98 0.98 0.98 0.98

P 06 1.35 1.35 1.35 1.35 1.35

P 07 1.24 1.25 1.31 1.32 1.32

P 08 3.19 3.34 3.47 3.48 3.48

P 09 2.44 2.70 3.62 2.97 3.03

P 10 2.62 2.45 2.69 2.71 2.72

4.2.3 Results Iteration Size 24 Hours with

Pheromones’ Limitations

To avoid the effect described above, we limited the

amount of pheromones that a user is allowed to spend

within an iteration. We obtained this by adding the

arithmetic mean of the pheromone’s values instead of

adding the values from each session (limited adjust-

ment). With this method we got stable NOPs (Tables

6 and 7).

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

684

However, there remains the challenge to interpret

the pheromone’s strength in the proﬁles, because the

values are still not normalised.

Table 6: NOPs average values / limited (24 h).

avg f=0.8 f=0.9 f=0.99 f=-0.002 f=-2.5*10

-4

P01 0.14 0.15 0.18 0.16 0.18

P02 0.05 0.05 0.06 0.05 0.06

P03 0.07 0.07 0.09 0.08 0.09

P04 0.06 0.06 0.08 0.07 0.09

P05 0.04 0.05 0.06 0.05 0.06

P06 0.09 0.10 0.12 0.11 0.12

P07 0.10 0.10 0.12 0.11 0.12

P08 0.29 0.30 0.34 0.32 0.35

P09 0.36 0.38 0.44 0.41 0.45

P10 0.29 0.31 0.37 0.34 0.38

Table 7: NOPs max. values / limited (24 h).

max f=0.8 f=0.9 f=0.99 f=-0.002 f=-2.5*10

-4

P01 1.34 1.50 1.72 1.71 1.74

P02 0.49 0.49 0.49 0.49 0.49

P03 0.92 0.92 0.92 0.92 0.92

P04 0.88 0.88 0.88 0.88 0.88

P05 0.95 0.95 0.95 0.95 0.95

P06 0.95 0.95 0.95 0.95 0.95

P07 0.95 0.95 0.95 0.95 0.98

P08 1.35 1.41 1.56 1.47 1.64

P09 1.23 1.34 1.90 1.53 1.90

P10 1.43 1.80 2.57 2.30 2.64

4.2.4 Environment for Predicting the Next Page

The primary goal of the algorithm is to lead visitors to

the content they are interested in. Because only log-

ﬁles were available, we did a ﬁrst test by measuring

how accurate we can predict visitors’ behaviour based

on the interest values we calculated.

We performed two series of experiments. In the

ﬁrst serie we split the original set of 6000 sessions

into the training and test sets which contained 4,000

and 2,000 sessions respectively. In the second serie

we split it into 5000 and 1000 sessions. To avoid a

high inﬂuence by short sessions, we only took ses-

sion with 3 ≤ session length ≤ 50. This reduced the

number to 2219 and 3013 sessions in the training sets.

We used these sessions to simulate visits on the

web-site. During the runs we calculated the visitors’

current proﬁle of interest at every 5th, 7th, and 9th

step (page). We ranked the outgoing links of the page

in order to compare these proﬁles with the links the

visitors follow during the session. This approach was

chosen to proof the assumption that the knowledge of

the visitor’s interests will help to get good results To

get a point of reference (benchmark), we compared

these results to a basic counting algorithm. We ini-

tialised the graph like in the previous approach. But

instead of calculating pheromone values, we counted

for each link how often it was taken by the visitors.

The obtained numbers were used during the simula-

tion to rank the outgoing links and predict the next

page. Table 8 gives an overview of the parameters for

this test.

Table 8: Testing environment parameters.

Parameter Value

# Sessions 6 000

# Topics 10

Iteration Size 24 h

# Training Sessions (4000) 2 419

# Training Sessions (5000) 3 013

4.2.5 Test Results for Predicting the Next Page

In the following, we present the results for the dif-

ferent runs. Tables 9 and 10 show the results for the

smaller training set while tables 11 and 12 presents

the results for the larger one. The rank values indicate

the rate of success based on the number of predictions

for the next page. For example: “Rank 2, 5th Page”

in table 9 indicates that we were successful in nearly

50% of all cases to predict the 6th page in a session,

if we are allowed to give two suggestions.

Table 9: Results (ANOP / Training Session (4000)).

ANOP. 5th Page 7th Page 9th Page

Rank1 336 (40.2%) 279 (41.3%) 248 (43.6%)

Rank2 419 (50.1%) 345 (51.0%) 302 (53.1%)

Rank3 484 (57.9%) 398 (58.9%) 340 (59.8%)

Rank4 521 (62.2%) 433 (64.0%) 375 (65.9%)

Rank5 556 (66.4%) 460 (68.0%) 402 (70.7%)

... ... ... ...

ALL 836 (100%) 676 (100%) 569 (100%)

Both algorithms provide relatively good results. A

larger set of training data leads, as expected, to a bet-

ter result in the veriﬁcation process. But it seems that

we can better predict the next pages by following the

majority of the previous visitors.

What are the conclusions from these results? On

the one hand it seems that the algorithm is not suit-

able to predict a visitor’s next. On the other hand, in

a productive system the algorithm would be used to

suggest a number of pages that will ﬁt in the best way

to a proﬁle. This can be every page on a web-site and

is not limited to the next reachable pages. It could

also be that the property of the algorithm to ’forget’

information after some time is turning in this case into

a disadvantage. As mentioned above, the results de-

pend on many parameters. To examine this behaviour

in a deeper way will be part of our future work.

FINDING NON-OBVIOUS PROFILES BY USING ANT-ALGORITHMS

685

Table 10: Results (Count / Training Session (4000)).

Count 5th Page 7th Page 9th Page

Rank1 368 (44.0%) 308 (45.6%) 265 (46.6%)

Rank2 457 (54.7%) 377 (55.8%) 329 (57.8%)

Rank3 513 (61.4%) 424 (62.7%) 371 (65.2%)

Rank4 562 (67.2%) 457 (67.6%) 401 (70.8%)

Rank5 601 (71.9%) 497 (73.5%) 420 (73.8%)

... ... ... ...

ALL 836 (100%) 676 (100%) 569 (100%)

Table 11: Results (ANOP / Training Session (5000)).

ANOP. 5th Page 7th Page 9th Page

Rank1 115 (41.3%) 139 (45.9%) 121 (47.1%)

Rank2 194 (51.7%) 168 (55.5%) 147 (57.2%)

Rank3 225 (60.0%) 190 (62.7%) 163 (63.4%)

Rank4 236 (62.9%) 201 (66.3%) 173 (67.3%)

Rank5 251 (66.9%) 213 (70.3%) 180 (70.4%)

... ... ... ...

ALL 375 (100%) 303 (100%) 257 (100%)

Table 12: Results (Count / Training Session (5000)).

Count 5th Page 7th Page 9th Page

Rank1 183 (48.8%) 156 (51.5%) 130 (50.6%)

Rank2 242 (64.5%) 196 (64.7%) 162 (63.0%)

Rank3 272 (72.5%) 223 (73.6%) 180 (70.0%)

Rank4 298 (79.5%) 236 (77.8%) 194 (74.5%)

Rank5 313 (83.5%) 253 (83.5%) 206 (80.2%)

... ... ... ...

ALL 375 (100%) 303 (100%) 257 (100%)

Both algorithms provide relatively good results. A

larger set of training data leads, as expected, to a bet-

ter result in the veriﬁcation process. But it seems that

we can better predict the next pages by following the

majority of the previous visitors.

What are the conclusions from these results? On

the one hand it seems that the algorithm is not suit-

able to predict a visitor’s next. On the other hand, in

a productive system the algorithm would be used to

suggest a number of pages that will ﬁt in the best way

to a proﬁle. This can be every page on a web-site and

is not limited to the next reachable pages. It could

also be that the property of the algorithm to ’forget’

information after some time is turning in this case into

a disadvantage. As mentioned above, the results de-

pend on many parameters. To examine this behaviour

in a deeper way will be part of our future work.

5 CONCLUSIONS

We have shown how the combination of the NOP and

ant-algorithm approach can help to derive individual

proﬁles once the topics-weights assignment and the

possibility to recognise users are taken into account.

Therefore it can be used to support the individual user.

Our current experiments are based on log ﬁles and our

subjective assignment of topics and their weights to

every page.

The most important future work will be to im-

prove the algorithm to produce normalised values for

a better comparison of the different models. It is also

important to implement feedback functionality and a

recommendation system that supports the visitor in an

active way. We would also like to automate the pro-

cess of topic assignment to make it less subjective.

ACKNOWLEDGEMENTS

This work was funded by the Fonds National de la

Recherche Luxembourg (FNR). We thank Prof. C.

Schommer (Universit

e du Luxembourg), Prof. R. Zi-

cari (Goethe-University, Frankfurt/Main, Germany)

and Nortel GmbH.

REFERENCES

Ambrosi, T. (2007). Erstellung und Analyse von nicht of-

fensichtlichen Proﬁlen mit Hilfe von Ameisenalgo-

rithmen.

Bonabeau, E., Dorigo, M., and Theraulaz, G. (1999).

Swarm Intelligence – From Natural to Artiﬁcal Sys-

tems. Oxford.

Cooley, R., Mobasher, B., and Srivastava, J. (1999). Data

preparation for mining world wide web browsing

patterns. In Knowledge and Information Systems.

Springer-Verlag.

Goss, S., Aron, J., Deneubourg, J., and Pasteelsi, J. (1989).

Self-organized shortcuts in the argentine ant. In

Naturwissenschaften 76, pages 579–581. Springer-

Verlag.

Hoebel, N., Kaufmann, S., Tolle, K., and Zicari, R. (2006).

The design of gugubarra 2.0: A tool for building and

managing proﬁles of web users. In Web Intelligence,

2006. WI 2006. IEEE/WIC/ACM International Con-

ference on.

Jeanson, R., Ratnieks, F. L. W., and Deneubourg, J.-L.

(2003). Pheromone trail decay rates on different sub-

strates in the pharaohs ant, Monomorum pharaonis.

In Physiological Entomology, volume 28, pages 192–

198.

Mushtaq, N., Tolle, K., Werner, P., and Zicari, R. (2004).

Building and evaluating nonobvious user proﬁles for

visitors of web sites. In e-Commerce Technology,

2004. CEC 2004. Proceedings. IEEE International

Conference on.

Traniello, J. F. A. (1989). Foraging strategies of ants. In

Annual Reviews Entomology, volume 34, pages 191–

210.

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

686