Everybody has a sense of trusting people or institutions, but how is trust de ned? It always depends on the speci c eld of research and application and is di erent most of the time, which makes it hard to answer this question in general at a computational level. Thinking on knowledge processing systems we have this question twice. How can we de ne and calculate trust values for the input data and, much more challenging, what is the trust value of the output? Meeting this challenge we rst investigate appropriate ways of de ning trust. Within this paper we consider three di erent existing trust models and a self developed one. Then we show ways, how knowledge processing systems can handle these trust values and propagate them through a network of processing steps in a way that the nal results are representative. Therefore we show the propagation of trust with the three existing trust models and with a recently self developed approach, where also precision- and importancevalues are considered. With these models, we can give insights to the topic of de ning and propagating trust in knowledge processing systems.
The main subject of our research is the topic of how knowledge processing systems can work with trust values. While going deeper into this research eld, several questions arise.
In our work, we try to gure out, how trust can be de ned and measured and in particular the question of how knowledge processing systems can deal with trust? Furthermore we investigate the topic of how several trust values can be combined (in general and in knowledge processing) and how can trust values be propagated through several steps of a knowledge processing system?
We try to give answers to these questions by investigating possibilities of trust measurement, combination and propagation and try to propose a sound and all-encompassing way of handling these topics.
The rest of this paper is structured as following: section 2 de nes common terminologies and shows related work in our eld of research.
As the term "trust" has a signi cant high importance in our research, we dedicated an extra section for de ning trust in knowledge processing - section 3. In this section, we investigate di erent models of measuring or determining trust. We brie y introduce our recently developed and already published approach, where trust- and precision values are handled, processed and propagated by taking into account several importances in section 4.
Section 5 examines the question, how trust can be propagated through knowledge processing systems. Here we cover the di erent trust measuring models, which were introduced in section 3 and 4 and how trust can be propagated in these di erent models. Section 6 shows the application of the presented trust propagation models in a scenario. We close this paper with section 7 by giving a summary of our work and an outlook for further research. 2
In this section we provide some insights into important terms which are relevant to our work.Furthermore we discuss the fusion of sensor precision values. 2.1
Trust
The meaning of the term "Trust" always depends on the speci c environment
and eld of research and application. In a recent publication about trust, we
state: "The question of 'How can we trust anything/anybody?' is discussed since
the beginning of mankind, but what does this topic mean in context to today's
technology age and especially for the information technology?" [
The three main types of applicable trust by Rousseau et al. [
Another point of view is the similarity of trusting people and trusting
technology, especially information technology, where the main di erence is within
the application of trust in the speci c area [
Also a very interesting publication about trust in information sources is from
Hertzum et al. [
Provenance
When we come into trust concerning trusting in data and trusting the sources
of data, the term "Data Provenance" comes into account. It means the origin
and complete processing history of any kind of data. A quite good introduction
and overview can be found in [
Several problems concerning data provenance are covered in [
Recent research work on provenance can be found in the following literature:
"Trust Evaluation Scheme of Web Data Based on Provenance in Social Semantic
Web Environments" [
"Research of Data Resource Description Method oriented Provenance" [
Risk Risk in general addresses the potential of losing something with a special personal value. It is also seen as an intentional interaction with uncertainty, where the outcome is hard to predict.
Rousseau et al. [
Precision & Multi Data Sensor Fusion
The link on related work of fusion precision values in sensor networks can be
found in our recent publication "Focussing on Precision- and Trust-Propagation
in Knowledge Processing Systems" [
The main question in our work is, how knowledge processing systems can handle and work with trust. In this context, the rst step is to nd a de nition of trust, which is suitable for this scienti c domain. In this section we investigate di erent applicable models of measuring or determining trust. Therefore we describe three existing ways: the binary trust model, the probabilistic trust model and the opinion-space trust model. 3.1
Trust, Certainty and Precision in Knowledge Processing
To the best knowledge of the authors, there is no related work dealing with this
topic directly { neither for processing trust and certainty, nor for the aggregation
of trust, (un)certainty or precision. A good approach for measuring trust is given
in [
The propagation/fusion of (sensor) precision values has been evaluated in recent publications, as stated in section 2.4. Some of the investigated propagation/fusion methods of sensor precision values seem quite promising also for the application on trust. Nevertheless we focus on models that cover only trust values, as presented in the following sections.
Another approach is the refactoring of a trust value from a given precision, which will be covered in our future work. 3.2
Binary Trust Model One of the easiest ways to represent trust values in an understandable and applicable way is the usage of a binary trust model. In this model the possible trust values can either be 0 or 1. Therefore the only di erentiation is to fully trust a subject (trust = 1) or not (trust = 0).
In our opinion, the model is very hard to apply in a real world domain because it is very hard to get a trust value of 1 anyway. The de nition of possible states in the binary trust model, can be seen in formula 1.
T = 0 _ 1 (1) Formula 1: Boundaries for the scope of trust in the binary trust model. 3.3
Probabilistic Trust Model A very application oriented and realistic way of representing trust is the usage of a probabilistic trust model. In this model the possible trust values range from 0 to 1 and can, for example be seen as a type of percentage view. The value of not trusting a subject (trust = 0) and fully trusting a subject (trust = 1 or 100%) is like in the binary trust model, but here the grading can be more precise, as there are (theoretically) in nite states of trust between 0 and 1 (or 100%). The range of possible states of trust in the probabilistic trust model can be seen in formula 2.
0
T 1 (2) Formula 2: Boundaries for the scope of trust in the probabilistic trust model. 3.4
Opinion-Space Trust Model
A very well developed model for measuring trust values is the opinion-space
model by Audun J sang and S.J. Knapskog from 1998: "A Metric for Trusted
Systems" [
They introduce an evidence space and an opinion space which are two equivalent models for representing human beliefs, which can be summarized as trust in their model. We focus on the opinion-space trust model, which consists of the values belief b, disbelief d, and uncertainty u. These three values represent the trust, which is determined. The sum of the three values is always 1, so the interpretation of trust has to be clari ed for the current domain of application. The opinion-space trust model can be seen in gure 1.
b + d + u = 1; fb; d; ug 2 [0; 1]3 (3) Formula 3: Boundaries for the scope of trust in the opinion-space model. 4
In our recent research, we designed a convenient approach for propagating trust
and precision values through multi-step knowledge processing systems, where
also a factor importance was introduced and considered in the calculation. Our
approach was evaluated and published in several conferences before e.g. in [
Principle Idea Following, we describe the idea of our approach. The main components are: { any Source (S), which provides data; there can be multiple sources. { any Data (D), which is provided by one source; for our model, every source usually provides one or more data (elements). { any Knowledge Processing System (KPS), which processes data from one or more sources; each KPS itself produces new data as output.
The main values in our approach are: { Trust value (T) of source (S), which de nes how trustable the source is. The system has to be seen as a whole environment, hence the trust level for one source should always be the same. { Precision value (P) of data (D), which describes how precise, reliable, con dent or steady the provided data is. { Importance value (I) of one input data (D), decided by the current knowledge processing system (KPS) for the current step of computation.
Our approach is sketched in gure 2. { Trust T of source S, for each S, has to be greater than 0 and less or equal than 1, where each value of T for each S has to be the same (if used multiple times) - a higher value represents higher trust: 0 < T 0 < P 1 1 { Precision P of data D, for each D, has to be greater than 0 and less or equal than 1, where each value of P for each D has to be the same (if used multiple times) - a higher value represents higher precision: { Importance I of data D, decided by the knowledge processing system: 0.5 for values which are not very important 1.0 for regular values, where no special impact on importance is given 1.5 for very important values, concerning the current data processing (4) (5) I = 0.5 j 1 j 1.5 (6)
In this section we compare di erent methods of fusing and propagating trust values in knowledge processing systems. 5.1
Propagating Trust in the Binary Trust Model As simple the determination of trust in the binary trust model is, as simple is the propagation of of these values. The only way of a reasonable propagation of binary trust values, is to take the results of the logical conjunction over all values. This means, if only one input value is 0, the output is 0 as well. This can be explained, because it is not justi able to trust an output of a processing where one input value wasn't trusted. (7) (8) (9) n Tnew = ^ Ti i=1 n Tnew = Y Ti i=1 Formula 8: Calculating Tnew over all T1-n in the probabilistic trust model. 5.3
Propagating Trust in the Opinion-Space Trust Model In this model, a reasonable way of propagation is to consider all input values in the opinion-space trust model. This means that the values of belief, disbelief and uncertainty have to be propagated through the system in a way that the output can again be identi ed with separate belief, disbelief and uncertainty values again. We propose a very clear way of propagating the three values through a knowledge processing system: using the arithmetic mean for all three separated values to gain the new output values.
Formula 7: Calculating Tnew over all T1-n in the binary trust model. 5.2
Propagating Trust in the Probabilistic Trust Model In the probabilistic trust model, several trust values can be fused by multiplying them, presupposed the providing sources are independent.
bnewjdnewjunew = n1 Xn (bijdijui) i=1 Formula 9: Calculating bdunew over all bdu1 n in the opinion-space trust model.
Propagating Trust in the Arithmetic Mean Trust Model This model of propagation is based on our approach, which is presented in section 4. It calculates the propagated trust value based on the input values from the sources, weighted with di erent important values, which are decided by the knowledge processing system itself individually for every step of processing. The scenario is a complex environment and consists of several knowledge processing systems, where multiple processing steps are taken and where some output values are used as new input values. We introduce six sources (S1 to S6) with di erent trust values (T1 to T6), each providing one or two data with di erent precision values for multiple knowledge processing systems (KPSA to KPSE), which weight the di erent importances. All following calculated propagation models consider the trust values. The usage of the precision- and importance values is only considered in the arithmetic mean trust model of our approach.
Propagation with the binary trust model Table 1 shows the initial trust values in the scenario for propagation with the binary trust model. By applying formula 7 in the calculations 14 and 15 from the propagation of trust in the binary trust model, the result is Tnew=0. Here we notice again the problem of the binary trust model, where only one untrusted input (T6) ruins the trust in the overall system.
Propagation with the probabilistic trust model Table 2 shows the initial
trust values in the scenario for propagation with the binary trust model.
By applying formula 8 in the calculations 19 and 20 from the propagation of
trust in the probabilistic trust model, the result is Tnew=0.07287 or 7.287%.
Again, we recognize the problem of inputs with "lower" trust values (T4 and
T5) which lead to very low overall trusting outcome, especially in multi step
knowledge processing systems.
Propagation with the opinion-space trust model Table 3 shows the initial
values for belief, disbelief and uncertainty in the scenario for propagation with
the opinion-space trust model.
By applying formula 9 in the calculations 30-35 from the propagation of trust in
the opinion-space trust model, the result is bnew=0:59166, dnew=0:23055, and
dnew=0:177. A very mean and realistic result in particular, when we observe the
well distributed input values. In our opinion, the outcome is very representative.
Propagation with the arithmetic mean trust model Table 4 shows the
initial trust values in the scenario for propagation with the weighted arithmetic
mean trust model.
The calculation steps for this scenario can be found in [
D11: P11=0.9 D2 : P2 =0.3 D31: P31=0.8 D32: P32=0.5 D4 : P4 =0.2 D51: P51=1.0 D52: P52=0.7 D6 : P6 =1.0
KPSA: IA1=0.5 KPSA: IA2=1.0 KPSA: IA3=1.5 KPSB: IB1=1.0 KPSB: IB2=1.5 KPSB: IB3=0.5 KPSP: IP1=1.0 KPSC: IC2=1.0 Because the scenario from the last subsection is completely arti cial, we applied our approach in the course of a project in the agricultural domain funded by the European Union, to have comparable computations.
Therefore, we are now referring to the DPM (Disease Pressure Model, used
in the Project CLAFIS [
We addressed the question of how to determine trust values in general and how knowledge processing systems can handle these values in particular. Additionally, we investigated several models for the de nition of trust and presented our recently developed approach. Furthermore we showed ways to propagate trust values from the investigated trust models through multi step knowledge processing systems and also a new way of trust propagation from our approach. We gave example calculations on a scenario with all propagation models and compared the results.
Our approach was evaluated in a real world scenario in the frame of a cooperation project with colleagues from practice. An evaluation was done by interviewing experts from this scienti c and application domain and the results were encouraging. Our aim is to develop a complete model for incorporating trust, precision and importance values into knowledge processing systems. This approach then could be applied to all other processing systems as well. Such a system, which could be used in a broad variety of applications, de nitively would be very useful in practice.
The research leading to these results, has received funding partly from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement no.604659 in the project CLAFIS and partly from the federal county of Upper Austria.