Twitter User Gender Inference Using Combined Analysis of Text and Image Processing

Abstract

Profile inference of SNS users is valuable for marketing, target advertisement, and opinion polls. Several studies examining profile inference have been reported to date. Although information of various types is included in SNS, most such studies only use text information. It is expected that incorporating information of other types into text classifiers can provide more accurate profile inference. As described in this paper, we propose combined method of text processing and image processing to improve gender inference accuracy. By applying the simple formula to combine two results derived from a text processor and an image processor, significantly increased accuracy was confirmed.

A Weighted Combination of Text and Image Classifiers for User Gender Inference

“…Ulges et al (2012) detected TV viewers' gender and age via content-based concept detection. Ikeda et al (2013) and Sakaki et al (2014) used methods that incorporate information. Ikeda et al (2013) proposed a hybrid-based method using both text and community membership.…”

“…Ikeda et al (2013) proposed a hybrid-based method using both text and community membership. Sakaki et al (2014) proposed a hybrid-based method using a combination of text and images, which builds a meta-classifier using the probability score out- put from text and image classifiers as input. This study demonstrated that a combination of text and images boosts the accuracy of a single source.…”

“…Tweet level annotation process refers to rules proposed by Sakaki et al (2014), who defined the tweet level labels as male and female. Workers annotated the labels using many sources of potentially discriminative metadata, including user preferences, icons, text, and images.…”

“…However, in many cases, SNS user information such as gender, age or residence is not openly available, although such information is extremely important for marketing. To meet that objective, several studies have been conducted to infer demographic information of anonymous users using text or image data posted on Twitter, and community membership (Rao and Yarowsky, 2010;Ikeda et al, 2013;Sakaki et al, 2014). Sakaki et al (2014) demonstrated that a hybridbased method outperformed other approaches using individual sources.…”

“…To meet that objective, several studies have been conducted to infer demographic information of anonymous users using text or image data posted on Twitter, and community membership (Rao and Yarowsky, 2010;Ikeda et al, 2013;Sakaki et al, 2014). Sakaki et al (2014) demonstrated that a hybridbased method outperformed other approaches using individual sources. However we observed an important issue: each probability score output from the image classifiers and the text classifier was simply summed, although the degree of their respective contributions to the inference is presumably different.…”