Phase 1-Identifying referring tasks based on category assessment32

categories. The referring tasks are then identified by calculating the similarity measures based on the relevance degrees of tasks to categories.

Step 1 of phase 1: Determine the semantic term set and corresponding fuzzy number

For modeling the workers’ perceptions on Relevance, the system defines six linguistic terms from “very low” to “perfect” to represent different relevance degrees.

Each worker has his/her own perception of the approximate value (fuzzy scale) for each linguistic term. The fuzzy scale of a linguistic term is often modeled as a triangular fuzzy number. The linguistic terms are displayed in the front-end interface to provide knowledge workers a more natural and easier way of relevance assessment, while the corresponding fuzzy number is in the back-end for the system to facilitate numerical computation of relevance ratings. Notably, evaluators may not have identical fuzzy numbers on six linguistic terms of “Relevance” owing to different perceptions of the linguistic terms. Table 2 lists six linguistic scales of

corresponding fuzzy numbers determined by four evaluators. For example, the evaluator E1’s perception on the linguistic term “very high” is within the fuzzy scale of (0.6, 0.7, 0.8). The evaluator E2’s perception on the linguistic term “very high” is within the fuzzy scale of (0.6, 0.75, 0.9). Each evaluator can use the front-end interface to easily setup his/her own fuzzy number of each linguistic term, or simply use the default fuzzy number provided by the system

A linguistic variable, Relevance, is defined to represent the degree of relevance between items (tasks or categories) assessed by evaluators. E(Relevance) is characterized using a fuzzy set of a universe of discourse U=[0,1], in which six linguistic terms řj and their associative semantic meanings m(řj) are defined as follows.

E(Relevance) = { ř0 = very low, ř1 = low, ř2 = normal, ř3 = high, ř4 = very high, ř5 = perfect}

where m(ři) < m(řj), for i < j, and all m(řj) are distributed in [0,1].

The anti-symmetric distributed term set [34] is adopted, where more positive linguistic terms are defined, as shown in the defined term set, since this work places more emphasis on positive feedback to items. The fuzzy linguistic approach models the meaning of each term by fuzzy numbers. This work employs triangular fuzzy number (TFN), as defined in Definition II of Appendix A-1, to express the fuzzy scale of each linguistic term. TFN is widely used owing to its simplicity and solid theoretical basis [59], and thus is used to represent each linguistic term of the

“Relevance“ variable.

Step 2 of phase 1: Assess the relevance of task to categories collaboratively

This step mainly assesses the relevance of the executing-task to each category. The executor, namely the knowledge worker with the executing-task at hand, rates the relevance of executing-tasks to each category by linguistic terms (e.g. low, high etc.).

Linguistic ratings denote the rating given in linguistic terms for the remainder of this paper. In addition, task-relevant experts or colleagues can also rate the relevance of executing-tasks to each category by linguistic terms to achieve collaborative

Table

2.

Corresponding fuzzy numbers of linguistic term set by different evaluators

Evaluators VL (Very Low) L (Low) N (Normal) H (High) VH (VeryHigh) P (Perfect) E1 (0,0.2,0.4) (0.3,0.4,0.5) (0.4,0.5,0.6) (0.5,0.6,0.7) (0.6,0.7,0.8) (0.7,0.8,0.9) E2 (0,0.1,0.2) (0.1,0.3,0.5) (0.4,0.5,0.6) (0.5,0.6,0.7) (0.6,0.75,0.9) (0.7,0.9,1) E₃ (0,0,0) (0.1,0.25,0.4) (0.3,0.4,0.5) (0.6,0.7,0.8) (0.7,0.8,0.9) (0.8,0.9,1) E4 (0,0,0) (0.1,0.3,0.5) (0.5,0.6,0.7) (0.6,0.7,0.8) (0.7,0.8,0.9) (0.8,0.9,1)

assessment. Collaborative assessment, where a comprised rating is derived by aggregating ratings of task-relevant experts or colleagues is especially useful for the executor who is unfamiliar with the executing task. The linguistic ratings cannot be used by the system to calculate aggregate ratings, and thus need to be transformed into crisp ratings. In the front-end, linguistic terms are used. In the back-end, the system transforms the linguistic ratings into crisp ratings. Four evaluators determine the degree of relevance of the executing task te to each category using linguistic ratings based on their subjective judgments, as listed in Table 3(A). The corresponding fuzzy number of each linguistic rating is transformed into crisp numbers (ratings). For example, evaluator E1’s perception of the linguistic term

“very high” is within the fuzzy scale of (0.6, 0.7, 0.8). The fuzzy number is transformed into a crisp value, 0.7, as shown in Table 3(B).

The fuzzy linguistic approach models the meaning of each term using fuzzy numbers. To achieve computational advantage, the crisp ratings (Best Non-fuzzy Performance values; BNP) are extracted from fuzzy numbers. Various methods can be used to defuzzify fuzzy numbers, including mean of maximal (MOM), center of area (COA), bisector of area (BOA), and so on [38].

This work adopts the COA method to calculate fuzzy numbers, owing to its simplicity and practicability. The COA method calculates the fuzzy mean under uniform probability distribution assumption [45]. If the fuzzy number Z



_is

triangular, the crisp rating can be derived by the equation:

( ) [( ) ( )] / 3

CV Z



= r l− + m l− +l. For example, Table 3(B) lists the crisp ratings transformed from the linguistic ratings of the evaluators based on the above equation.

Step 3 of phase 1: Aggregate the relevance ratings of evaluators

Evaluators’ crisp ratings obtained from collaborative assessment are aggregated in this step. The relevance degree of the executing task to each category is derived by computing the weighted average of evaluators’ crisp ratings of

Table

3

. Assess the relevance of executing task to categories

(A) Assessment by linguistic terms (B) Crisp ratings derived from linguistic ratings

Evaluator Evaluator

Category

E1 E2 E3 E4

Categor

y E1 E2 E3 E4

C1 N N N H C1 0.5 0.5 0.4 0.7

C2 VH H N VH C2 0.7 0.6 0.4 0.8

C3 P VH VH H C3 0.8 0.75 0.8 0.7

C4 VL N L L C4 0.2 0.5 0.25 0.3

relevance to the category. The aggregated relevance of the executing task to categories is expressed as a vector of relevance degree to each category. The top-N similar tasks and last-M non-similar tasks can then be retrieved based on the similarity measures between the relevance degrees of tasks to categories, as detailed in Step 4. Let Aej

(c

i

) denote the crisp rating of evaluator e

j regarding the relevance of the executing task te to category ci. Moreover, let wej denote the associated weight representing the relative importance (weight) of the rating of evaluator ej. The aggregated relevance of the executing task to category ci

, A

E(ci), is derived as ∑j number of evaluators, then the aggregated relevance ratings are calculated as the arithmetic mean. The aggregated relevance degrees of the executing task to categories are expressed as follows.

1 2 3 4 0.525, 0.625, 0.7625, 0.3125 ( ), ( ), ( ), ( )

C

e E E E E

tJK =< A c A c A c A c >=< >

Step 4 of phase 1: Select referring tasks

The proposed mechanism reduces the number of tasks to be assessed via discovering a set of referring tasks to assist workers in conducting task-relevance assessment (phase 2 assessment). This step identifies a subset of existing tasks as referring tasks based on their similarity to the executing task derived using the relevance degrees of tasks to categories. Notably, relevance degrees of the executing task to categories are derived by Step 3 of the category assessment. A similarity (cosine) measure is adopted to calculate the similarity between the executing task and an existing task based on their relevance degrees to categories. Based on the similarity measures, the top-N similar tasks are chosen as the positive (relevant) referring tasks, whereas the last-M non-similar tasks are chosen as the negative (irrelevant) referring tasks. The referring tasks are used for further task-relevance assessment in phase 2.

The similarity measure between the executing task te and an existing task tr

can

be computed as the cosine of the angle between two vectors,

t

JJK_e^C

and

t

JJK_r^C

, namely

cosine( t

JJK_e^C

,

t

JJK_r^C

). Notably,

t

JJK_e^C

is derived by the collaborative relevance assessment as

described in Step 3, while

t

JJK_r^C

is derived by the fuzzy classification, as described in Chapter 4.

Example: The aggregated relevance degrees of the executing task te to categories is modeled as a vector _tJK_e^C

,

t

^JJK_e^C =< 0.525, 0.625, 0.7625, 0.3125 > . Table 4 lists the relevance of ten existing tasks to categories. The similarly measures of executing task and existing tasks are listed in the sixth column of Table 4. The ranking of similarity measures is displayed in the last column. The top-5 tasks, t3, t4, t5, t9 and t10, are selected as the positive referring tasks, while the last-2 tasks, t2 and t6, are chosen as the negative referring tasks.

5.3.2 Phase2-Assessing the relevance of referring tasks

Phase 2 conducts an assessment to determine the relevance of the referring tasks to the executing task. The evaluators assess the degree of relevance between the executing task and referring tasks without reviewing all tasks. The task assessment procedure resembles the procedure of category assessment. The evaluators conduct relevance assessment to determine the relevance degree of each referring task to the executing task. They use linguistic terms to rate the relevance of each referring task to the executing task. The aggregated relevance rating of a referring task is derived by computing the weighted average of evaluators’ crisp ratings on the relevance of the referring task to the executing task. The relevance degrees of referring tasks to the executing task are then used to construct the task profile of the executing task detailed in Section 5.4.

Let

A

ej

(t

r

) represent the crisp rating of the evaluator e

j on the relevance of the executing-task to a referring task tr. Moreover, let wej denote the associated weight

Table

4.

Relevant degree between tasks and categories

C1 C2 C3 C4 Similarity Ranking

Task 1 0 0.25 0.22 0 0.838 (6)

Task 2 0 0 0 0.71 0.269 (10)

Task 3 0 0.25 0.24 0 0.844 (5)

Task 4 0.12 0.19 0.29 0 0.947 (3)

Task 5 0 0.11 0.13 0 0.850 (4)

Task 6 0 0 0.68 0 0.657 (9)

Task 7 0.49 0.11 0.15 0 0.724 (8)

Task 8 0 0.15 0.54 0 0.778 (7)

Task 9 0.18 0.18 0.29 0 0.957 (1)

representing the relative importance (weight) of the rating of evaluator ej. The aggregated relevance rating of task tr to the executing-task, AE(tr), is derived as

∑j

w

ej

A

ej

(t

r

).

Example: Table 5(A) lists the linguistic ratings on five positive referring tasks evaluated by four evaluators. Moreover, Table 5 (B) lists the crisp ratings of tasks 5 and 9, derived from the linguistic ratings. The aggregated relevance ratings,

A

E(t5)=0.675 and AE(t9)=0.8375, are also listed in the last column of Table 5 (B).

Table

5.

Assessment on the relevance of positive referring tasks to the executing task (A) Assessment by linguistic terms

Evaluator Referring Task

E1 E2 E3 E4

Task 3 VL VL H VL

Task 4 VL VL VL VL

Task 5 H H VH H

Task 9 P VH P P

Task 10 VL H H H

(B) Crisp ratings derived from linguistic ratings Evaluator Referring Task

E1 E2 E3 E4

Aggregated relevance rating of tasks AE(tr)

Task 5 0.6 0.6 0.8 0.7 0.675

Task 9 0.8 0.75 0.9 0.9 0.8375

5.3.3 Discussions

In this work, the task-related experts of each task are predefined. In addition, the relative importance of experts is given the same weight to aggregate the relevance ratings. That is, the aggregated relevance ratings are calculated as the arithmetic mean. In the future, we shall consider revising our group decision method with the aid of recommendation techniques in Recommender system. Accordingly,

y We will employ methods e.g., collaborative filtering algorithm, demographic profiles of workers, etc, to determine task-related experts. Thus, the new-system cold-start problem may encounter by the demographic profiles of workers and the new-user cold-start problem my encounter by the hybrid recommendation technique.

y Furthermore, the relative importance of task-related experts could be determined by the calculation result of recommendation algorithms.

在文檔中 以工作觀為基礎之知識支援模式與系統:工作相關知識遞送與分享 (頁 46-52)