PROVIDING RECOMMENDATIONS IN AN AGENT-BASED
TRANSPORTATION TRANSACTIONS MANAGEMENT
PLATFORM
Alexis Lazanas, Nikos Karacapilidis and Yiannis Pirovolakis
Industrial Management and Information Systems Lab, MEAD
University of Patras, 26504 Rio Patras, Greece
Keywords: Recommendation techniques, recommender systems, transportation, web-based systems, intelligent agents.
Abstract: Diverse recommendation techniques have been already proposed and encapsulated into several e-business
systems aiming to perform a more accurate evaluation of the existing alternatives and accordingly augment
the assistance provided to the users involved. Extending previous work, this paper focuses on the
development of a recommendation module for transportation transactions purposes and its integration in a
web-based platform. The module is built according to a hybrid recommendation technique, which combines
the advantages of collaborative filtering and knowledge-based recommendations. The proposed technique
and supporting module enable customers to consider in detail alternative transportation transactions
satisfying their requests, as well as to evaluate such transactions after their completion.
1 INTRODUCTION
Transportation management involves diverse
decision making issues, which are basically related
to the choice of route and carrier selection. Such
issues raise due to the differentiation between the
dispatcher’s preferences (e.g. cost limitation, loading,
and delivery dates) and the carrier’s service
resources (e.g. transportation media, itinerary, and
capacity). The matching of the above preferences
and services cannot be easily handled manually, as
in most cases a plethora of alternative options exist,
while time and money limitations are ubiquitous.
Therefore, the field of transportation management
requires quick and cost-effective solutions to the
customers’ demands for distribution and shipping
operations.
This paper extends our previous work on the
exploitation of software agent technology in
transportation management (Karacapilidis et al.,
2006; Lazanas et al., 2005). More specifically, we
have addressed analysis, design and implementation
issues raised during the development of an
innovative agent-mediated electronic marketplace,
which is able to efficiently handle transportation
transactions of various types. Agents of the proposed
system represent and act for any user involved in a
transportation scenario, such as customers who look
for efficient ways to ship their products and
transport companies that may - fully or partially -
carry out such requests, while they cooperate and get
the related information in real-time mode. Our
overall approach is based on flexible models that
achieve efficient communication among all parties
involved, coordinate the overall process, construct
possible alternative solutions and perform the
required decision-making. In addition, the
supporting web-based system is able to handle the
complexity that is inherent in such environments,
which is mainly due to the frequent need of finding a
“modular transportation solution”. To further
explain this concept, consider the case where a
customer wants to convey some goods from place A
to place B, while there is no transport company
acting directly between these two places. Supposing
that two available carriers X and Y have some
scheduled itineraries from A to C and from C to B,
respectively, it is obvious that a possible solution to
the above customer’s request is to involve both X
and Y and fragment the intended overall itinerary to
the related sub-routes. It is also noted that these
carriers may be associated with diverse
transportation means, such as trains, trucks, ships
and airplanes. Our system is able to manage all the
necessary freighting and fleet scheduling processes
in wide-area transportation networks. Its agents
87
Lazanas A., Karacapilidis N. and Pirovolakis Y. (2006).
PROVIDING RECOMMENDATIONS IN AN AGENT-BASED TRANSPORTATION TRANSACTIONS MANAGEMENT PLATFORM.
In Proceedings of the Eighth International Conference on Enterprise Information Systems - SAIC, pages 87-92
DOI: 10.5220/0002461900870092
Copyright
c
SciTePress
cooperate upon well-specified business models,
which may efficiently carry out diverse processes.
This paper focuses on the features and
functionalities of a new module integrated in the
above system, namely the recommendation module,
which aims at enhancing the quality of the
associated decisions. Recommender systems have
been described as systems that produce
individualized recommendations as output or have
the effect of guiding the user in a personalized way,
in environments where the amount of on-line
information vastly outstrips any individual’s
capability to survey it (Burke, 2002). Alternative
techniques have been also proposed in the literature
in order to handle the above issues. Having
thoroughly considered their pros and cons, our
approach follows a hybrid recommendation
technique.
The remainder of this paper is structured as
follows. Section 2 discusses approaches and related
work from the area of recommender systems.
Section 3 describes the basic aspects of our
approach, namely transportation plan selection,
alternative solution evaluation and recommendation
methodology. Section 4 deals with the evaluation of
a transaction and the exploitation of agent
technologies. Finally, Section 5 concludes the paper.
2 RECOMMENDER SYSTEMS
Recommender systems apply data analysis
techniques to assist users finding the items they need
by producing a predicted likeness score or a list of
top-N recommended items. Two of the most widely
adopted recommendation techniques are
Collaborative Filtering (CF) and Knowledge Based
Recommendation (KBR), each one possessing its
own strengths and weaknesses. Collaborative
Filtering (CF) (Resnick et al., 1994) is the most
commonly used recommendation technique to date.
The basic idea of CF-based algorithms is to provide
item recommendations or predictions, based on the
opinion of other like-minded users. In a typical CF
scenario, there is a list of m users U = {u
1
,u
2
,…,u
m
}
and a list of n items I = {i
1
,i
2
,…,i
n
}. Each user u
i
is
associated with a list of items I
Ui
, for which the user
has expressed his/her opinion. Opinions can be
explicitly given by the user as a rating score,
generally within a certain numerical scale, or can be
implicitly derived from transaction records, by
analyzing timing logs, mining web hyperlinks and so
on. For a particular user U
a
, the task of a
collaborative filtering algorithm is to find an item
likeness that can be of two forms:
• Prediction: this is a numerical value, P
a,j
,
expressing the predicted likeness of item i
j
(i
j
does not belong in I
Ua
) for the user. The
predicted value is within the same scale
(e.g. 1 to 5) as the opinion values provided
by U
a
.
• Recommendation: this is a list of N items I
r
(I
r
is a subset of I) that the user will like
most (the recommended list must contain
items not already selected by the user). This
outcome of CF algorithms is also known as
Top-N recommendation. (Sarwar et al.,
2000).
On the other hand, Knowledge-Based
Recommendation attempts to suggest objects based
on inferences about a user’s needs and preferences.
In some sense, all recommendation techniques could
be described as doing some kind of inference.
Knowledge-based approaches are distinguished in
that they utilize functional knowledge: that is, they
have knowledge about how a particular item meets a
particular user need, and can therefore reason about
the relationship between a need and a possible
recommendation. The user profile can be any
knowledge structure that supports this inference. In
the simplest case, as in Google (www.google.com),
it may simply be the query that the user has
formulated. The Entrée system and several other
recent systems (for an overview, see (Schmitt &
Bergmann, 1999)) employ techniques from case-
based reasoning for knowledge-based
recommendations.
The knowledge used by a knowledge-based
recommender system can also take many forms.
Google uses information about the links between
web pages to infer popularity and authoritative value
(Brin & Page, 1998). Entrée uses knowledge of
cuisines to infer similarity between restaurants.
Utility-based approaches calculate a utility value for
objects to be recommended; in principle, such
calculations could be based on functional
knowledge. However, existing systems do not use
such inference mechanisms, thus requiring users to
do their own mapping between their needs and the
features of products, either in the form of preference
functions for each feature, as in the case of Tête-à-
Tête, or answers to a detailed questionnaire, as in the
case of PersonaLogic (Burke, 2002). Knowledge-
based recommender systems are prone to the
drawback of all knowledge-based systems: the need
for knowledge acquisition. More specifically, there
ICEIS 2006 - SOFTWARE AGENTS AND INTERNET COMPUTING
88
are three types of knowledge that are involved in
such a system:
• Catalog knowledge: Knowledge about the
objects being recommended and their
features. For example, the Entrée
recommender should know that “Thai”
cuisine is a kind of “Asian” cuisine.
• Functional knowledge: The system must be
able to match the user’s needs with the
object that might satisfy those needs. For
example, Entrée knows that a need for a
romantic dinner spot could be met by a
restaurant that is “quiet with an ocean
view”.
• User knowledge: To provide good
recommendations, the system must have
some knowledge about the user. This might
take the form of general demographic
information or specific information about
the need for which a recommendation is
sought.
Of these knowledge types, the last one is the
most challenging, as it is, in the worst case, an
instance of the general user-modeling problem.
Despite this drawback, knowledge-based
recommendation has some beneficial characteristics.
It is appropriate for casual exploration, because it
demands less from the user compared to the utility-
based recommendation. Moreover, it does not
involve a start-up period during which its
suggestions are of low quality. On the other hand, a
knowledge-based recommender cannot “discover”
user niches, the way collaborative systems can.
However, it can make recommendations as wide-
ranging as its knowledge base allows.
3 OUR APPROACH
3.1 Transportation Plans and
Evaluation of Alternative
Solutions
The recommendation procedure adopted in our
approach is highly associated with the selection (by
the user) of the appropriate “transportation plan”
(see Figure 1). A transportation plan typically
defines the user preferences for the upcoming
transactions. The five alternative plans offered are:
• Express
• Economic
• Safe
• Dependable
• User Defined
It can be easily observed that each of the first
four plans declares a specific tension in the
recommendation strategy to be followed by the
system, in that it either minimizes the overall
duration or cost (first two plans) or it retains a high
level of safety or dependability (third and fourth
plans) of the suggested itineraries. The last choice
offers the possibility for a customized plan
definition. A user defined plan may combine
parameters from all the above four plans. The
selection of one of these plans will influence the
recommendations processes of our approach for the
particular user.
Figure 1: User request for a transaction.
Having constructed all possible transportation
solutions, our approach proceeds to the evaluation
phase. In this phase, our approach’s Decision_Maker
agent performs two distinct tasks. First, a refinement
of the set of alternative solutions constructed takes
place, by excluding solutions that do not comply
with the customer’s requirements (the algorithm
developed for the construction of alternative routes
takes into account only the duration and cost
parameters as upper bounds of the proposed
solutions). In this phase, a set of predefined rules are
employed to exclude the alternative solutions that do
not correspond to the specific freight
transportation’s requirements and customer
preferences. For instance, if the customer had chosen
the “Safe” transportation plan, the following set of
rules will be deployed:
for each constructed_solution
{
if (safety_level < AVERAGE) OR
((safety_level = AVERAGE) AND
PROVIDING RECOMMENDATIONS IN AN AGENT-BASED TRANSPORTATION TRANSACTIONS
MANAGEMENT PLATFORM
89
(dependability_level < LOW)) then
discard constructed_solution;
else if (safety_level >= HIGH) OR
((safety_level = AVERAGE) AND
(dependability_level >= LOW)) then
solution ( constructed_solution;
}
Table 1 presents the constraints to be met for
each transportation plan (for the User Defined plan,
this process takes into account the constraints set by
the user). In all cases, solutions that do not satisfy
these constraints are discarded.
Table 1: Selection Criteria for each plan.
Plan
1
Cost Duration Safety
Dependa-
bility
1 Any Min Any Any
2 Min Any Any Any
3 Any Any >Average >=Low
4 Any Any >= Low > Average
5 UD
2
UD UD UD
1
Plans 1 to 5 correspond to: Express, Economic, Safe,
Dependable and Hybrid.
2
User Defined Criterion
3.2 A Methodology for the Selection
of Alternative Route Paths
In order to construct optimal and sub-optimal
solutions, our approach uses an elaborated version of
Dijkstra’s shortest path algorithm (Crauser et al.,
1998). The majority of shortest path algorithms in
the literature uses a bidirectional, single-weighted
graph to represent a connected set of vertices (V
i
)
through a number of arcs A
ij
(from V
i
to V
j
). Our
algorithm takes into consideration each A
ij
and its
correspondent weight (w
ij
) in order to produce a
route path from a starting point (S) to an ending
point (E) that minimizes the total weight (w
SE
). The
complexity in our approach consists in the presence
of a pair of factors that affect each arc’s weight,
namely the cost and the duration. Due to the fact that
there exist two weights for each arc, we confronted
the problem of unifying these weights into a single
one, in order to proceed with the ranking of the
solutions.
More specifically, each arc’s A
ij
weight (W
ij
)
consists of a cost weight (
W
cost-ij
) and a duration
weight (
W
duration-ij
). Obviously, W
ij
= W
cost-ij
+
W
duration-ij
.
Having defined the total weight for each arc we
encountered the problem of adding these two
parameters that are measured in different units
(Euros and hours, respectively). This problem was
confronted by applying a normalization technique
that divided both the cost and duration terms of an
A
ij
route with the correspondent maximum cost and
duration. It is:
w
duration-ij
= duration
ij
/max(duration)
w
cost-ij
= cost
ij
/max(cost)
Another issue that came up after the weight
normalization procedure concerned the solutions
ranking. To address this problem, our approach
provides the user with different solutions by using a
pair of weight coefficients (
costCoef and
durationCoef) and calculating solutions
corresponding to alternative combinations of the
weights of the cost and duration criteria (see Figure
2), according to the formula:
w
ij
=(costCoef * w
cost-ij
) +
(durationCoef * w
duration-ij
)
Figure 2: Coefficients’ variation.
This process is described in pseudo-code below:
{
costCoef Å 0.0;
durationCoef Å 1.0;
step ( 0.0;
for each step calculate
{
weight[i][j] ( costCoef*Wcost +
durationCoef * Wduration;
costCoef ( step;
durationCoef = 1-step;
}
perform shortest path algorithm;
step ( step + 0.1;
}
The outcome of this process is then provided to
the user. An instance of the related system interface
is shown in Figure 3.
ICEIS 2006 - SOFTWARE AGENTS AND INTERNET COMPUTING
90
3.3 Recommendation Methodology
The recommendation procedure begins immediately
after the ranking of the alternatives. It is a complex
process which is carried out in three basic stages: the
evaluation of the carriers and the transactions data,
the exploitation of transaction data through a data
mining process, and the recommendation
methodology selection or synthesis. At the
beginning of the process, the system stores all the
appropriate data that are submitted by the user and
are related with scheduled or completed transactions.
These data are of significant importance and will be
further exploited by the data mining process.
Moreover, in this stage the user evaluates the
carrier(s) involved in a transaction through an
interface since the evaluation process concerns
assigning a score to each carrier (this is discussed in
more detail in Section 4.1).
Figure 3: Solution ranking produced by the system.
The second stage of recommendation concerns
the data mining process. At this stage, transactions
data are gathered through knowledge construction
processes (Cho et al., 2002). Although important,
the description of these processes goes out of the
scope of this paper. We only mention here that
knowledge construction in our case refers to the
formulation of rules that could be used when a
knowledge based recommendation strategy is
performed.
The last phase of the recommendation refers to
the selection or synthesis of the appropriate
recommendation technique. This objective will be
reached through the definition of well structured
rules that will be applied for each transaction. The
recommender agent of our system takes the initiative
to look for the most appropriate recommendation
technique. For example, for a particular itinerary
from point i to j, taking into consideration that the
customer has selected a certain plan, a rule for the
specific itinerary could recommend a carrier that is
different than the one suggested by the CF
technique, based on the carriers’ evaluation process
described earlier in this section.
4 IMPLEMENTATION ISSUES
The recommender module has been integrated into
our platform as an add-on tool and has been
thoroughly tested. After the completion of a
transportation transaction, the customer is able to
evaluate it by providing his/her opinion about the
carrier(s) who were involved. Since transactions
handled by our platform often concern modular
solutions, multiple carriers may be involved. Thus,
the rating procedure concerns each individual
carrier.
Figure 4: The carriers’ evaluation interface.
As shown in Figure 4, a customer expresses
his/her position for a completed transaction by
selecting a value for each feature, following a typical
Likert scale (very low, low, average, high, very
high). Each value is stored in the appropriate table in
the systems SQL Server database. This rating is very
helpful during the main phase of the
recommendation process presented in Figure 5,
which takes place after the ranking of the alternative
solutions described earlier. As shown in the figure,
for the itinerary “London–Prague” the customer has
selected one of the proposed routes provided by the
system (see Figure 3). After selecting the specific
route, the customer may request a recommendation
for the involved carriers. The system provides
different types of recommendations features such as:
PROVIDING RECOMMENDATIONS IN AN AGENT-BASED TRANSPORTATION TRANSACTIONS
MANAGEMENT PLATFORM
91
an average score ranking, top-10 carriers, “red card”
carrier (carriers with under average scoring), and a
“suggestion” functionality. Each suggestion made by
the system is accompanied by a prediction status
label P
i
= {critical, acceptable, indifferent}, which
refers to the significance level of the system’s
suggestion. In the example shown in Figure 5, the
suggestion is concerned “critical” because the
system retains some “negative” knowledge for the
carrier under consideration.
Figure 5: The recommendation module of the system.
For the implementation of the proposed
recommendation technique, a recommender module
has been developed using agent and data
warehousing technologies. More specifically, a new
agent, namely the Recommender Agent, has been
implemented in order to coordinate the overall
process of recommendation and has been assigned
with two main tasks:
• The recommendation technique selection,
in that the agent is capable of switching
among alternative recommendation
algorithms, depending on the characteristics
of each transaction. The agent carries out
data mining tasks and generates the
appropriate rules or models.
• The characterization of each carrier who is
involved in a transportation transaction and
the recommendation of alternative carriers
capable of fulfilling a particular
transportation request.
Through these formally modeled tasks, the
recommender agent provides continuous assistance
to customers, while it remains active and capable to
adapt its “behavior” into a rapidly changing
environment.
5 CONCLUSIONS
This paper presents our approach on the integration
of hybrid recommendation techniques into an agent–
based transportations transaction management
platform. We proposed a hybrid approach that
combines different recommendation techniques, in
order to provide the user with more accurate
suggestions. The overall process is coordinated by a
recommender agent, who is responsible of
performing multiple tasks. The presence of the agent
guarantees that the user will be provided with
continuous recommendation, dynamic update and
recommendation techniques synthesis.
REFERENCES
Brin, S. and Page, L.: 1998, The anatomy of a large-scale
hyper textual {Web} search engine. Computer
Networks and ISDN Systems, 30(1-7), 107-117.
Burke, R. (2002). Hybrid Recommender Systems: Survey
and Experiments. User Modeling and User-Adapted
Interaction, 12, pp. 331-370.
Cho, Y., Kim J. and Kim, S. (2002). A personalized
recommender system based on web usage mining and
decision tree induction. Expert Systems with
Applications, 23, pp. 329-342.
Crauser A., Mehlhorn K., Meyer U. and Sanders P., 1998.
A Parallelization of Dijkstra's Shortest Path
Algorithm. In Proceedings of 23rd International
Symposium, MFCS'98, Brno, Czech Republic.
Karacapilidis N., Lazanas A., Megalokonomos G. and
Moraitis P., 2006: On the Development of a Web-
based System for Transportation Services. Information
Sciences (in press).
Lazanas A., Evangelou C., and Karacapilidis N. (2005):
Ontology-Driven Decision Making in Transportation
Transactions Management. In W. Abramowicz (ed.),
Proc. of the 8th Int. Conf. on Business Information
Systems, Poznan, Poland, April 20-22, pp. 228-241.
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and
Riedl, J. (1994). GroupLens: An Open Architecture
for Collaborative Filtering of Netnews. In Proceedings
of CSCW '94, Chapel Hill, NC.
Sarwar, B.M., Karypis, G., Konstan, J.A. and Riedl, J.
(2000). Analysis of Recommender Algorithms for E-
Commerce. In Proceedings of the ACM Conference
on e-Commerce, Minneapolis, MI, October 17-20,
2000, ACM Press, pp. 158-167.
Schmitt, S. and Bergmann, R.: 1999, Applying case-based
reasoning technology for product selection and
customization in electronic commerce environments.
12th Bled Electronic Commerce Conference. Bled,
Slovenia, June 7-9, 1999.
ICEIS 2006 - SOFTWARE AGENTS AND INTERNET COMPUTING
92
