Change search
ReferencesLink to record
Permanent link

Direct link
A Continuous Dataflow Pipeline For Low Latency Recommendations
KTH, School of Information and Communication Technology (ICT).
2016 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
Abstract [en]

The goal of building recommender system is to generate personalized recommendations to users. Recommender system has great value in multiple business verticals like video on demand, news, advertising and retailing. In order to recommend to each individual, large number of personal preference data need to be collected and processed. Processing big data usually takes long time. The long delays from data entered system to results being generated makes recommender systems can only benefit returning users. This project is an attempt to build a recommender system as service with low latency, to make it applicable for more scenarios. In this paper, different recommendation algorithms, distributed computing frameworks are studied and compared to identify the most suitable design. Experiment results reviled the logarithmical relationship between recommendation quality and training data size in collaborative filtering. By applying the finding, a low latency recommendation workflow is achieved by reduce training data size and create parallel computing partitions with minimal cost of prediction quality. In this project the calculation time is successfully limited in 3 seconds (instead of 25 in control value) while maintaining 90% of the prediction quality.

Place, publisher, year, edition, pages
TRITA-ICT-EX, 2016:6
URN: urn:nbn:se:kth:diva-180695OAI: diva2:896151
Available from: 2016-01-20 Created: 2016-01-20 Last updated: 2016-01-20Bibliographically approved

Open Access in DiVA

fulltext(5255 kB)105 downloads
File information
File name FULLTEXT01.pdfFile size 5255 kBChecksum SHA-512
Type fulltextMimetype application/pdf

By organisation
School of Information and Communication Technology (ICT)

Search outside of DiVA

GoogleGoogle Scholar
Total: 105 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

Total: 661 hits
ReferencesLink to record
Permanent link

Direct link