EXIST 2024

Tags

Introduction

Score prediction task

Given known subject-item preferences, predict new subject-item preferences. Formally, let $U$ a set of all subjects $U$ and $V$ a set of all items, our core task it is to find a real-valued scalar function $score(u,v)$ where $u\in U$ and $v \in V$ . To provide a hard label or multi labels, $k$ subjects vote with their encoded scores. Hence, we’ve reduced our problem into a score prediction problem.

For instance, the below table illustrates a complete utility matrix with known score entries $f(u,v)$ where 0 represents the label “NO” and 1 represents the label “YES”. In this case, the voting policy selects the class annotated by more than 3 subjects.

	$V_1$	$V_2$	$V_3$
$U_1$	1	0	1
$U_2$	0	1	1
$U_3$	1	0	0
$U_4$	1	1	1
$U_5$	0	0	0
$U_6$	1	0	0
Voting	1	0	Undefined
Label	YES	NO

There are different methods such as memory-based CF, model-based CF, Neural CF, GCN-based CF.

Task 2.

$score(u,v)=0 \implies$ NO

$score(u,v)=1 \implies$ Direct

$score(u,v)=2 \implies$ Judgemental

Task 3.

$score(u,v)=0b000001 \implies$ -

$score(u,v)=0b000010 \implies$ {IDEOLOGICAL-INEQUALITY}

$score(u,v)=0b000100 \implies$ {MISOGYNY-NON-SEXUAL-VIOLENCE}

$score(u,v)=0b00010|0b00001=0b00011\implies$ {-, IDEOLOGICAL-INEQUALITY}

Embedding-based models

We consider embedding-based models for scoring. For each user $u\in U$ , let $u \in \mathbb{R^D}$ for its $D$ -dimensional embedding. For each item $v\in V$ , let $v \in \mathbb{R}^D$ be its $D$ -dimensional embedding. So, $score(u,v) \equiv f: \mathbb{R}^D \times \mathbb{R}^D \to \mathbb{R}$ .