Details
| Originalsprache | Englisch |
|---|---|
| Titel des Sammelwerks | SAC '19 |
| Untertitel | Proceedings of the 34th ACM/SIGAPP Symposium on Applied Computing |
| Seiten | 564-573 |
| Seitenumfang | 10 |
| Publikationsstatus | Veröffentlicht - 8 Apr. 2019 |
| Veranstaltung | 34th Annual ACM Symposium on Applied Computing, SAC 2019 - Limassol, Zypern Dauer: 8 Apr. 2019 → 12 Apr. 2019 |
Abstract
Opinion stream classification algorithms adapt the model to the arriving review texts and, depending on the forgetting scheme, reduce the contribution old reviews have upon the model. Reviews are assumed independent, and information on the entity to which a review refers, i.e. to the opinion target, is thereby ignored. This implies that the prediction of a review's label is based more on reviews referring to other, more popular or simply more recently inspected entities, while reviews referring to the same entity might be ignored as too old. In this study, we enforce that the reviews to each entity are taken into account for learning, adaption and forgetting. We split the original stream to substreams, each substream comprised by the reviews referring to the same entity (opinion target). This allows us to deal with differences in the speed of each substream and to exploit the impact of the entity itself on the labels of the reviews referring to it. For this constellation of substreams we propose a pair of two voting classifiers, one being the global, “entity-ignorant” classifier trained on the whole stream of reviews, the other one consisting of one “entity-centric” classifier per entity. We show that the entity-ignorant classifier contributes most for entities with very few reviews, i.e. during the cold-start, while the entity-centric classifiers contribute most after acquiring enough information on the corresponding entities. We study our approach on a stream of product reviews, show that our ensemble improves the performance of its members, and we discuss the conditions under which one member contributes more than the other.
ASJC Scopus Sachgebiete
- Informatik (insg.)
- Software
Zitieren
- Standard
- Harvard
- Apa
- Vancouver
- BibTex
- RIS
SAC '19: Proceedings of the 34th ACM/SIGAPP Symposium on Applied Computing. 2019. S. 564-573.
Publikation: Beitrag in Buch/Bericht/Sammelwerk/Konferenzband › Aufsatz in Konferenzband › Forschung › Peer-Review
}
TY - GEN
T1 - Exploiting Entity Information for Stream Classification over a Stream of Reviews
AU - Beyer, Christian
AU - Matuszyk, Pawel
AU - Unnikrishnan, Vishnu
AU - Ntoutsi, Eirini
AU - Niemann, Uli
AU - Spiliopoulou, Myra
N1 - Funding Information: This work is partially funded by the German Research Foundation, project OSCAR "Opinion Stream Classification with Ensembles and Active Learners". Additionally, the first author is also partially funded by a PhD grant from the federal state of Saxony-Anhalt.
PY - 2019/4/8
Y1 - 2019/4/8
N2 - Opinion stream classification algorithms adapt the model to the arriving review texts and, depending on the forgetting scheme, reduce the contribution old reviews have upon the model. Reviews are assumed independent, and information on the entity to which a review refers, i.e. to the opinion target, is thereby ignored. This implies that the prediction of a review's label is based more on reviews referring to other, more popular or simply more recently inspected entities, while reviews referring to the same entity might be ignored as too old. In this study, we enforce that the reviews to each entity are taken into account for learning, adaption and forgetting. We split the original stream to substreams, each substream comprised by the reviews referring to the same entity (opinion target). This allows us to deal with differences in the speed of each substream and to exploit the impact of the entity itself on the labels of the reviews referring to it. For this constellation of substreams we propose a pair of two voting classifiers, one being the global, “entity-ignorant” classifier trained on the whole stream of reviews, the other one consisting of one “entity-centric” classifier per entity. We show that the entity-ignorant classifier contributes most for entities with very few reviews, i.e. during the cold-start, while the entity-centric classifiers contribute most after acquiring enough information on the corresponding entities. We study our approach on a stream of product reviews, show that our ensemble improves the performance of its members, and we discuss the conditions under which one member contributes more than the other.
AB - Opinion stream classification algorithms adapt the model to the arriving review texts and, depending on the forgetting scheme, reduce the contribution old reviews have upon the model. Reviews are assumed independent, and information on the entity to which a review refers, i.e. to the opinion target, is thereby ignored. This implies that the prediction of a review's label is based more on reviews referring to other, more popular or simply more recently inspected entities, while reviews referring to the same entity might be ignored as too old. In this study, we enforce that the reviews to each entity are taken into account for learning, adaption and forgetting. We split the original stream to substreams, each substream comprised by the reviews referring to the same entity (opinion target). This allows us to deal with differences in the speed of each substream and to exploit the impact of the entity itself on the labels of the reviews referring to it. For this constellation of substreams we propose a pair of two voting classifiers, one being the global, “entity-ignorant” classifier trained on the whole stream of reviews, the other one consisting of one “entity-centric” classifier per entity. We show that the entity-ignorant classifier contributes most for entities with very few reviews, i.e. during the cold-start, while the entity-centric classifiers contribute most after acquiring enough information on the corresponding entities. We study our approach on a stream of product reviews, show that our ensemble improves the performance of its members, and we discuss the conditions under which one member contributes more than the other.
KW - Document Prediction
KW - Entity-Centric Learning
KW - Stream Classification
UR - http://www.scopus.com/inward/record.url?scp=85065637268&partnerID=8YFLogxK
U2 - 10.1145/3297280.3297333
DO - 10.1145/3297280.3297333
M3 - Conference contribution
AN - SCOPUS:85065637268
SN - 9781450359337
SP - 564
EP - 573
BT - SAC '19
T2 - 34th Annual ACM Symposium on Applied Computing, SAC 2019
Y2 - 8 April 2019 through 12 April 2019
ER -