Grbovic, M., Radosavlijevic, V., Djuric, N., Bhamidipati, N., & Nagarajan, A. (2016). Leveraging Blogging Activity on Tumblr to Infer Demographics and Interests of Users for Advertising Purposes. In #Microposts2016 Workshop Proceedings.
As one of the leading platforms for creative content, Tumblr offers advertisers a unique way of creating brand identity. Advertisers can tell their story through images, animation, text, music, video and more, and they can promote that content by sponsoring it to appear as an advertisement in the streams of Tumblr users. In this paper, we present a framework that enabled one of the key targeted advertising components for Tumblr, specifically, gender and interest targeting. We describe the main challenges involved in the development of the framework, which include the creation of a ground truth for training gender prediction models, as well as mapping Tumblr content to an interest taxonomy. For purposes of inferring user interests, we propose a novel semi-supervised neural language model for categorization of Tumblr content (i.e., post tags and post keywords). The model was trained on a large-scale data set consisting of 6.8 billion user posts, with a very limited amount of categorized keywords, and was shown to have superior performance over the baseline models. We successfully deployed gender and interest targeting capability in Yahoo production systems, delivering inference for users that covers more than 90% of the daily activities on Tumblr. Online performance results indicate advantages of the proposed approach, where we observed a 20% increase in user engagement with sponsored posts in comparison to untargeted campaigns.
Identifying the Dataset
Tumblr is a blogging and social network platform that hosts rich multimedia content and a dynamic follower-following ecosystem. Like many other online platforms, Tumblr does not require users to disclose their gender information when signing up. To obtain this data, this paper proposes leveraging descriptive blog data to infer gender. Sourced from the Tumblr Firehose database, the training dataset consisted of 6.8 billion posts, 96.9 million users, and 5.1 billion edges (follows) from which keywords were extracted and tags/html/keywords were discarded. For ground-truth gender mapping, researchers used the US census data of popular baby names from 1880-2013.
This study contributes to advertising, marketing, and microtargeting efforts. More specifically, the methods seek to enhance the effects of native advertising - advertising which resembles native content - by specifying both demographic and interest categories. The authors suggest a semi-supervised neural language model to combine and embed learnings related to keywords, tags, and categories simultaneously. Based on user profiles and online behaviour, historical ad targeting methods (solely relying on demographic attributes) can expand to include interests. Interest-taxonomy models were optimised via scholastic gradient ascent for large-scale data applicability. In addition to interest predictions, the gender classification model tested in this paper used logistic regression based on user names and profile descriptions. Gender attributes are determined by thresholding probabilities and are estimated as linear inputs.
Key Assumptions Stated by Authors
Challenges posed by the Tumblr infrastructure and privacy concerns were among the key assumptions and limitations discussed. Firstly, while modeling, the distinct content and language features on Tumblr must be considered. Secondly, the vast amounts of unlabeled data and minimal labeled data (acquired by human efforts) makes it difficult to determine user interest categories.
With regard to privacy, the authors are constrained in terms of what kind of data can be used. User searches, clicks, and page visits are considered sensitive data and are not publicly available.
Despite these limitations, the author argues that personalisation via ad targeting would improve user experience and increase advertiser revenues.
This paper is unique in its focus on interest categorisation (despite operating in tandem with gendered linear regressions). Diverging away from relying solely on a gendered binary, the users are perceived more holistically. However, the intention of identifying and monitoring user trends does not align with the initial motivations users themselves have for creating their own profile. Hence, leveraging user profiles for more specific advertising and assuming that native advertising improves the user’s experience is questionable. Moreover, concerns about privacy stop at data availability and neglect to address the consent of the user to 1) have their data used for this study and 2) to be the target of ads.
7400 labeled tags were tested using the semi-supervised model from the 1000 editorially labeled tags used to train the model. The authors concluded that the classification of the remaining 7400 achieved higher precision and recall scores than other methods. For the gender experiment, Vowpal Wabbit and Hadoop were used to test 64.1 million users with 1007 users edited manually. Given the number of “not sure” cases that came up - gender is evidently difficult to infer, in comparison to the proposed interest-segmentation approach. Testing the validity of the results is ongoing as decayed categories and counts are updated on a daily basis to factor in the newest activities.
These results show the general futility of gender inference efforts and the success of interest taxonomies. Although, the results do little to demonstrate any impacts on advertising and or the user experience as an implication of better user categorisation.