Kogukonnapõhine aktiivsuse languse ennustamine sotsiaalvõrgustikus
Failid
Kuupäev
2014
Autorid
Ajakirja pealkiri
Ajakirja ISSN
Köite pealkiri
Kirjastaja
Abstrakt
Virtuaalsete sotsiaalvõrgustike haldajate seisukohalt on oluline tuvastada kasutajaid, kes kaotavad suure tõenäosusega lähitulevikus huvi nende teenuse vastu. Selliste kasutajate ennustamine lubab suunata neile kampaaniaid, hoidmaks või suurendamaks aktiivsust võrgustikus. Nimetatud probleemi lahendatakse sageli masinõppemeetodite abil, tehes ennustusi üksikisiku tasandil. Olemasolevad lahendused ei kasuta aga maksimaalselt ära kasutajate omavahelisi suhteid.
Selles kontekstis tutvustame uut lähenemist, ennustamaks aktiivsuse langust kogukondade ehk omavahel tihedalt seotud kasutajate gruppide tasandil. Antud töös kasutame kahte meetodit kogukondade leidmiseks ning võrdleme tulemusi üksikkasutajate ja juhuslike kasutajate gruppidega. Analüüs näitab, et teenusest loobuda plaanivaid kasutajaid on lihtsam leida kogukondade kui üksikisiku tasandil. Tulemused näitavad, et ennustuste kvaliteet sõltub ka kasutatud kogukondade leidmise algoritmist. Meetod, mis leiab kogukonnad lokaalsel tasandil, lähtudes iga kasutaja otsesest suhtlusringkonnast, võimaldab paremaid ennustusi kui võrgustikule tervikuna orienteeritud meetod. Lisaks eelmainitule võimaldab kogukonnapõhine analüüs arvesse võtta täiendavaid tunnuseid, saamaks täpsemaid ennustusi. Saadud tulemused on aluseks uute kogukonnapõhiste meetodite väljatöötamisele, analüüsimaks kasutajate aktiivsust sotsiaalvõrgustikes ning tõstmaks turunduskampaaniate efektiivsust.
An important problem for facilitators of online social networks is to identify the users who are likely to decrease their level of activity in the near future. Such predictions are the basis for targeted campaigns aimed at sustaining or increasing the overall user engagement in the network. A common approach to this problem is to apply machine learning methods to make predictions at the level of individual users. The existing approaches, however, do not consider the social connections of the individuals to their full extent, leaving room for improvement. In this context, we propose a new approach to the problem of activity decay prediction based on the idea of identifying groups of tightly inter-linked users (namely communities) where the level of social activity is likely to decay. We investigate two community detection methods and compare the resulting predictive accuracy against several baselines. We show that more individuals who are likely to decay can be reached by targeting communities instead of single users. Moreover, a bottom-up community detection method produces higher accuracy in this context than a top-down modularity-based approach. Additionally, a richer set of features related to user engagement can be used for prediction purposes, leading to more accurate predictions. The results pave the way for designing community-based approaches to analyze user engagement in social networks as well as associated community-based targeting methods.
An important problem for facilitators of online social networks is to identify the users who are likely to decrease their level of activity in the near future. Such predictions are the basis for targeted campaigns aimed at sustaining or increasing the overall user engagement in the network. A common approach to this problem is to apply machine learning methods to make predictions at the level of individual users. The existing approaches, however, do not consider the social connections of the individuals to their full extent, leaving room for improvement. In this context, we propose a new approach to the problem of activity decay prediction based on the idea of identifying groups of tightly inter-linked users (namely communities) where the level of social activity is likely to decay. We investigate two community detection methods and compare the resulting predictive accuracy against several baselines. We show that more individuals who are likely to decay can be reached by targeting communities instead of single users. Moreover, a bottom-up community detection method produces higher accuracy in this context than a top-down modularity-based approach. Additionally, a richer set of features related to user engagement can be used for prediction purposes, leading to more accurate predictions. The results pave the way for designing community-based approaches to analyze user engagement in social networks as well as associated community-based targeting methods.