TRACE: A Dynamic Model of Trust for People-Driven Service - - PowerPoint PPT Presentation

TRACE: A Dynamic Model of Trust for People-Driven Service Engagements

Combining Trust with Risk, Commitments, and Emotions Anup K. Kalia

Advisor: Munindar P . Singh Department of Computer Science North Carolina State University Raleigh, NC 27695, USA

September 30, 2015

Anup Kalia (NCSU) TRACE September 30, 2015 1 / 24

Broader Objectives

◮ Understand subtle human and organizational relationships ◮ Use such relationships as a basis for estimating trust

• Anup Kalia (NCSU)

TRACE September 30, 2015 2 / 24

Research Question

◮ How to estimate trust between people from their interactions?

Possible Applications

◮ Support people to make important decisions in organizational settings ◮ Estimating team cohesion or performance

Anup Kalia (NCSU) TRACE September 30, 2015 3 / 24

Limitations With Existing Approaches

Several approaches consider commitments alone for trust estimation.

◮ Gambetta (1988) interprets trust as a truster’s assessment of a trustee for

performing a specific task

◮ Mayer et al. (1995) define trust as the willingness of a truster to be

vulnerable to a trustee for the completion of a task

◮ Teacy et al. (2006) consider trust as the truster’s estimation of probability

that a truster will fulfill it’s obligation toward a trustee

◮ Wang et al. (2011) represent trust as the belief of a truster that trustee will

• cooperate. They estimate trust by aggregating positive and negative

experiences

◮ Kalia et al. (2014) consider commitment outcomes to predict trust where

they learn truster’s parameters based on whether outcomes are positive, negative, or neutral

Anup Kalia (NCSU) TRACE September 30, 2015 4 / 24

Limitations With Existing Approaches

Two major classes of trust models

◮ Fixed parameter trust models where parameter are manually fixed ◮ Machine-learned trust models typically Hidden Markov Models (HMM)

that assumes variables are conditionally independent of each other given the output variable

Anup Kalia (NCSU) TRACE September 30, 2015 5 / 24

Proposed Approach

We can improve trust prediction by incorporating (in addition to commitments) two attributes

◮ Risk taken by a truster toward a trustee

◮ Risk taken depends on a truster’s belief about the likelihood of gains or

losses it might incur from its relationships with a trustee

◮ Emotions displayed by a truster toward a trustee

◮ Studies in psychology suggest that positive emotions increase trust

whereas negative emotions decrease trust

◮ Create TRACE a model based on Conditional Random Field (CRF)

◮ Conditional independences between risk, commitments, and emotions may

not hold in our setting (e.g., in HMM)

Anup Kalia (NCSU) TRACE September 30, 2015 6 / 24

Background: Commitment & Trust

• Anup Kalia (NCSU)

TRACE September 30, 2015 7 / 24

Background: Commitment Lifecycle

◮ C(Debtor, Creditor, Antecedent, Consequent)

violated conditional discharged consequent detached cancel null create expire

• terminated

cancel

Anup Kalia (NCSU) TRACE September 30, 2015 8 / 24

Background: Estimating Trust from Commitment Progression

◮ Two-valued representation, positive and negative experiences: r, s ◮ Trust α =

r r+s

◮ We characterize each subject via four parameters

◮ Initial values, rin, sin ◮ Increment for positive and negative experiences: ir and is

Commitment Operation Trust rfi, sfi Commissive create

λir + rin, (1-λ)is + sin

Directive create Delegate None Discharge

ir + rin, sin

Cancel

rin, is + sin

Anup Kalia (NCSU) TRACE September 30, 2015 9 / 24

Trust Antecedent Framework (Mayer et al., 1995)

We propose TRACE based on the enhance trust antecedent framework

• ◮ The model contains 4 variables trust (T), risk (R), commitments (C), and

emotions (E)

◮ Each variable V = T, R, C, E is described using using Singh’s (1999,

2011) formal notation Vdebtor, creditor, antecedent, consequent

Anup Kalia (NCSU) TRACE September 30, 2015 10 / 24

Description of Variables

◮ Ctrustee, truster, antecedent, consequent

◮ The trustee commits to the truster to perform the consequent ◮ If the trustee performs the consequent, the commitment is satisfied

◮ Rtruster, trustee, antecedent, consequent

◮ The truster takes a risk by accepting the trustee’s offer to perform the

consequent

◮ If the trustee performs the consequent, the truster gains

◮ Ttruster, trustee, antecedent, consequent

◮ The truster believes the trustee if the trustee performs the consequent ◮ Trust has three dimensions: ability (trustee’s competency), benevolence

(trustee’s willingness), integrity (trustee’s ethics and morality)

◮ Etruster, trustee, antecedent, consequent

◮ The truster displays a positive emotion if the trustee performs the

consequent

Anup Kalia (NCSU) TRACE September 30, 2015 11 / 24

Postulates

We propose postulates that capture relationships between the variables P1 : Tt → Tt+1. The trust Tt+1 is influenced by the past trust Tt P2 : Ct → Tt. The current commitment outcome Ct influences the current trust Tt P3 : Rt → Ct. The risk taken influences the commitment outcome Ct or the gain or loss realized in the risk Rt P4 : Rt → Tt. The current risk taken Rt influences the current trust Tt P5 : Ct → Et. The commitment outcome Ct influences the current emotion Et P6 : Rt → Et. The risk taken Rt influences the truster’s emotion Et P7 : Et → Tt. The current emotion Et influences the current trust Tt

Anup Kalia (NCSU) TRACE September 30, 2015 12 / 24

The TRACE Model

Graphical representation of HMM and TRACE trust models (two time slices)

• Anup Kalia (NCSU)

TRACE September 30, 2015 13 / 24

Comparing HMM and CRF

◮ HMM makes two independent assumptions

◮ The current state yt is independent of y1, y2, . . ., yt−2, given yt−1 ◮ Observations xt are independent of each other, given yt

◮ CRF

◮ CRFs are agnostic to dependencies between the observations ◮ CRF model employs discriminative modeling, where the distribution p(

y| x) is learned directly from the data

Anup Kalia (NCSU) TRACE September 30, 2015 14 / 24

Evaluation

We evaluate TRACE via data collected from a human-subject study conducted by the Intelligence Advanced Research Projects Activity (IARPA)

◮ IARPA prepared a dataset based on the Checkmate protocol adapted

from the investment or dictator economic decision-making game (Berg, 1995)

◮ The data consists of 431 rows collected from 63 subjects ◮ Each row corresponds to the sequence of rounds played between two

subjects

◮ The data we obtained reflects only the banker’s perspective

• Anup Kalia (NCSU)

TRACE September 30, 2015 15 / 24

Evaluation

◮ The protocol involves two roles: banker and game player ◮ The banker’s task is to loan money to a game player ◮ The game player requests a loan from the banker to play a maze,

promising to play a maze of certain difficulty and return: the loan with all gains, the loan with 50% of all gains, 50% of the available money, a fixed amount

◮ After the game player’s request, the banker chooses a loan category:

small (1–7 USD), medium (4–10 USD), or big (7–13 USD)

◮ A dollar amount, randomly generated within the banker’s chosen

category, is loaned to the game player

◮ The game player does not know the category chosen by the banker ◮ The game player plays a maze of a certain difficulty (not necessarily what

he or she had promised)

◮ The banker will not know the actual maze played ◮ The game player returns some money to the banker (not necessarily what

he or she had promised)

Anup Kalia (NCSU) TRACE September 30, 2015 16 / 24

Mapping Game Elements to The TRACE Model

◮ The commitment from the player to the banker,

Cplayer,banker,loan,return, as satisfied if the player returned at least the amount he or she had loaned, and as violated, otherwise

◮ We compute the gain or loss in the risk, R banker, player, loan, return,

based on the difference between the loaned and returned amounts

◮ The dataset represents the banker’s trust for the player after the round,

Tbanker,player,loan,return, as a three-tuple A, B, I, indicating the banker’s perception of player’s ability, benevolence, and integrity

◮ The dataset represents the banker’s emotion after he or she receives a

return from the player, Ebanker,player,loan,return, as real-valued (1–10) state anxiety scores derived from the post-round questionnaire

Anup Kalia (NCSU) TRACE September 30, 2015 17 / 24

Results

MAEs of HMM and TRACE considering different feature combinations. Model Input Variables A B I HMM C 1.1220 0.8564 1.0917 C + R 0.8974 0.7655 0.8484 C + E 0.8619 0.7433 0.7184 R + E 0.8468 0.8376 0.7992 C + R + E 0.8870 0.7977 0.7714 TRACE C 0.8744 0.7576 0.7988 C + R 0.7463 0.7685 0.7876 C + E 0.8617 0.7656 0.7580 R + E 0.8949 0.6815 0.6568 C + R + E 0.7878 0.7427 0.7141

Anup Kalia (NCSU) TRACE September 30, 2015 18 / 24

Results

◮ Considering only C, TRACE yields lower MAEs than HMM for each trust

attribute (CRF employs discriminative)

◮ Considering all features (C + R + E), TRACE again yields lower MAEs

than HMM for each trust attribute (CRF captures dependencies between C, R, and E)

◮ Considering C and R, TRACE performs better than HMM in predicting A

and I (C and R are not independent)

◮ Considering C and E, HMM performs better than TRACE for B and I (C

and E are conditionally independent)

◮ Considering R and E, TRACE performs better than HMM for B and I

whereas HMM performs better than TRACE for A (mixed)

Anup Kalia (NCSU) TRACE September 30, 2015 19 / 24

Threats to Validity

◮ Our dataset, although real, consists of short sequences. We expect both

HMM and TRACE to perform better given longer sequences

◮ The dataset is skewed toward positive trust values and our conclusions

may not hold since the trust values have a different distribution

◮ The dataset represents emotions using anxiety scores only, thereby

lacking realistic emotion responses along multiple dimensions such as anger and joy

Anup Kalia (NCSU) TRACE September 30, 2015 20 / 24

Discussion

◮ TRACE illustrates that a probabilistic model of trust that incorporates

commitments, risk, and emotions can produce trust estimates with fairly good accuracy

◮ Our findings therefore open up the possibility of developing user agents

that promote secure collaboration

◮ Using TRACE a user can calibrate the perceived trust with the risk

undertaken in light of available measures of risk and gain from commitments

Anup Kalia (NCSU) TRACE September 30, 2015 21 / 24

Thanks

Anup Kalia (NCSU) TRACE September 30, 2015 22 / 24

Background: Identifying Commitment Operations from Interactions

Ten-fold cross-validation using SVM on marked up Enron email sentences

Commitment Operation Precision Recall F-measure Commissive create 0.87 0.97 0.92 Directive create 0.94 0.97 0.95 Delegate 0.86 0.33 0.48 Discharge 1 0.02 0.04 Cancel Features used in the classifier include (out of 15) 1 Modal verb (shall, will, may, might, can, could, would, must) 2 Type of subject (first person, second person, third person) 3 Present tense verb 4 Past tense verb 5 Deadline

Anup Kalia (NCSU) TRACE September 30, 2015 23 / 24

Background: Determining Commitment Operations and Trust from Text

Commitments being the most prominent normative relationship

S R Content Operation TS,R TR,S Kim Dorothy I will also check with Al- liance Travel Agency . . . create(C1) Kim Dorothy I checked with our Travel Agency . . . discharge(C1)

↑

Rob Kim By Wednesday Aug 16 2001, please send all copies of your documenta- tion . . . create(C2) Kim Rob Rob, please forgive me for not sending this in by Aug 15 cancel(C2)

↓

◮ Example emails from the Enron corpus

Anup Kalia (NCSU) TRACE September 30, 2015 24 / 24