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
H.2.8 [Database applications]: Data Mining data mining; computational advertising; audience modeling; algorithms Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citatAoPCioCnuonMlpbyoylnmifrsotiuhhgrseehtpdbfirtsretaihvpcsoaanpgt2oeear0.rea1tCdn6.oodApfhybatersichltgdraehadtcsb#teiyfmnMogriawcicucoirpttmhohuppcorororpen(sdoestisn)tset/2issso0.opw1fe6nrtmheWirist(twoesdr)ok;.rkTschoooowcpponypeipyndrogbotyhcpeeoreetwrhdmiesirnesi,tgtothsrea,drnepauvbaliislha,btloepoonstlionnesearsveCrsEoUrtRoreVdoislt-r1ib6u9te1t(ohltisttsp,:r/e/qcuieruesr-prwiso.rospregc/iVficople-rm16is9s1io)n and/or a fee. Request permissions from Permissions@acm.org.
K#DMD’i1c5ro,Apougsutst21001-61,3,A2p01r51,S1ythdn,e2y,0N16S,WM,Aounstrraélaial., Canada. c 2015 ACM. ISBN 978-1-4503-3664-2/15/08 ...$15.00.
DOI: http://dx.doi.org/10.1145/2766XXX.XXXXXXX.
In recent years, online social networks have evolved to
become an important part of life for online users of all
demographic and socio-economic backgrounds. They allow users
to easily stay in touch with their friends and family, discuss
everyday events, or share their interests with other users
with the click of a button. Tumblr is one such social
network, representing one of the most popular and fastest
growing networks on the web. Hundreds of millions of people
around the world come every month to Tumblr to find,
follow, and share what they love. The Tumblr network is a gold
mine of content, comprising of 200 million blogs on different
topics such as travel, sports, and music, where 85 million
user posts are published on a daily basis. This wealth of
user-generated data opens a great opportunity for
advertisers, allowing them to promote their products through
highquality targeting campaigns to both blog visitors and blog
owners [
The standard, prevalent form of advertising on Tumblr is
through sponsored posts that appear alongside regular posts
in the user’s dashboard, the central page for a Tumblr user,
displaying the newest posts of followed blogs in the form
of a stream. This form of advertising, in which
advertisements resemble native content in the stream, is often referred
to as native advertising. Native advertisements are usually
aesthetically beautiful and highly engaging, which typically
makes them more enjoyable than regular display ads [
Building of interest targeting products on social and
microblogging platforms is an important research topic,
discussed previously by several researchers [
which we explain in detail in this paper. In particular, the content and language used on Tumblr have distinct characteristics that needed to be accounted for during the modeling. Users often use tags to summarize the text in their posts. However, the language styles used in the tags and post text are different (e.g., the tag “hp” and the word “hp” have different meanings when they appear in the posts, “Harry Potter” and “HP company”, respectively). Moreover, unlike the popular social platform Facebook, which contains a large amount of social interactions but a limited amount of content, or the microblogging platform Twitter, which contains an intermediate amount of social interaction and content, Tumblr represents a unique combination of a rich and diverse content platform and a dynamic social network. To make use of this vast advertising potential, we propose to classify user-generated Tumblr content into a standard multi-level general-interest taxonomy 2 that advertisers commonly use for defining their targeting campaigns, opening doors to high-quality audience segmentation and modeling for purposes of ad targeting. However, inferring categories of user posts is a challenging task, given the huge quantities of unlabeled data being posted every day and the very limited amount of labeled data, typically obtained by human editorial efforts. To this end, we propose a novel semi-supervised neural language model, capable of jointly learning embeddings of post keywords, post tags and category representations in the same feature space. The neural model was trained on a large-scale data set comprising of 6.8 billion posts, with only a fraction of categorized content.
Targeting pipelines described in this paper are being used to show ads to millions of users daily, and have substantially improved Tumblr’s business metrics following the launch. On our path to developing targeting capabilities for Tumblr, we first created user profiles, based on users’ Tumblr activities that include publishing blog posts, following other blogs, liking posts, and other. Lastly, we aimed at building and delivering both demographic and interest predictive models based on the created profiles.
We note that the privacy of our users is of critical importance. Therefore, we were constrained in regards to what data we can use. Specifically, user profiles were created solely from data which users share publicly with others, including contents of blog posts, blog titles and descriptions, and follow, like and reblog actions. This data is publicly available through Tumblr Firehose data source3. Other user activities, such as user searches on Tumblr, which blogs they visited and where they clicked, are considered to be sensitive data and were not used in any way for the development of the ad targeting models. 2.
Personalization is defined as “the ability to proactively
tailor products and product purchasing experiences to tastes of
individual consumers based upon their personal and
preference information” [
We consider personalization through the domain of ad
targeting [
One of the basic approaches in ad targeting is to target
users with ads based on their demographics, such as age or
gender. Historically, this approach has proven to work better
than targeting random users. However, while for some
products this type of targeting may be sufficient (e.g., women’s
makeup, women’s clothing, man’s razors, man’s clothing),
for others it is not effective enough and a more involved
profiling of users is required. A popular method in today’s
ad targeting that addresses this issue is known as interest
targeting, in which users are assigned interest categories,
such as “sports” or “travel”, based on their historical
behavior [
targeting categories, and a model is learned to categorize
user activities and estimate their interest in each category.
Interest targeting is known to build good brand awareness
with relevant audience, which has already shown interest in
the corresponding category. In this paper we follow this
interest targeting approach. Alternatively, advertisers may be
interested going a step forward and optimizing for intent,
typically done by assigning categories to actual ads, and
training a machine learning model to estimate the
probability of an ad click in that category [
To the best of our knowledge, the Tumblr social network
has been considered by only a few scientific studies. In [
Tumblr5 is one of the most popular social blogging platforms, where users can create and share posts with the followers of their blogs. According to data from January 20156, there is a total of 221.6 million blogs on Tumblr, which jointly produced over 102.7 billion blog posts. With a large number of users signing up every day, Tumblr is currently the fastest growing social platform7. 3.1
To register for a Tumblr account, a valid e-mail address is required, along with a primary username (which will become a part of the blog URL) and a confirmation of age. A Tumblr blog resembles a webpage, with a profile picture, blog title and blog description appearing at the top (see Figure 1), followed by a stream of blog posts bellow. The first blog 5www.tumblr.com 6www.tumblr.com/about 7http://t.co/3txHFRJreJ title body tags reblog like created by a registered user is considered his or her primary blog. In addition, a very small portion of users maintain one or more secondary blogs. A Tumblr user is uniquely described by the blog ID of the primary blog, and throughout the paper we will use “blog” and “user” interchangeably.
Common user activities of Tumblr users include the
following: 1) creating a post on one’s blog; 2) sharing a post
created by another blog, called reblogging (a reblogged post
will appear on the user’s blog); 3) liking a post by another
blog; and 4) following another blog. Similar to Twitter, the
follow connections at Tumblr are uni-directional. However,
unlike Twitter, users can create longer and richer content in
the form of several post types, such as text, photo, quote,
link, chat, audio and video. The most popular types of blog
posts are photo posts and text posts, and, based on the
analysis published in [
Advertising on Tumblr is implemented through the mechanism of sponsored (or promoted) posts shown in a user’s dashboard. This is similar to how advertising works on Twitter and Facebook. A sponsored post can be a video, an image, or simply a textual post containing an advertising message. In Figure 4, we show an example of a sponsored post and how it appears on web and mobile dashboards. Similarly to organic (or non-promoted) posts, sponsored posts can propagate from user to user in the network by means of reblogs, and users can also “like” the promoted post. Both likes and reblogs can be seen as an explicit form of acceptance or endorsement of the advertising message. Moreover, THAT ARE FFECTIVE images by video rease the just like any other posts, sponsored posts are supplemented with notes on who liked and reblogged the post.
Interestingly, while user-generated, organic posts are reblogged on average 14 times, sponsored posts are reblogged on average 10,000 times8. We have observed that 40% of engagements with sponsored posts are reblogs, likes, or follows. What is more, every fourth reblog of a sponsored post results in 6 downstream reblogs from followers, leading to content longevity, and one third of reblogs of sponsored posts are present for 30 days or more after the initial post.
In this section, we describe the data sources (user activities and post contents) utilized to create user profiles. In particular, user activities included actions such as posts, likes, follows and reblogs, while post contents included tags, the title and body for text posts, artist names from audio posts, as well as tags and captions for photo posts. 4.1
Once signed in onto Tumblr, a user can follow other users’ blogs. The follow action is one-directional as it does not require the follow back. For the purpose of this study, we collected a sub-graph which contained 96.9 million unique nodes (i.e., users), 5.1 billion edges (i.e., follows), out of which 36.4 million are bi-directional (18.2 million pairs of users that follow each other). The data set included more than 26.1 billion activities on Tumblr. As discussed earlier, an activity log is available through a data feed called Firehose.
To create user profiles for targeting, textual contents of all posts were collected, including photo captions, tags, titles and bodies. In addition, every time a user performs a post or reblog activity, Firehose lists the user’s blog title and blog description, which we also used to represent a user. As we can see in Figure 1, a blog title and description often provide useful information with respect to targeting, such as
In order to improve the representation of user profiles, we propose to extract keywords from available blog information. This requires a certain amount of data preparation and processing. Given the extracted blog data, including the title, content and tags, we first removed all the html tags, followed by the removal of stopwords and the formation of bigrams. It is common for certain words to appear together more often than some others (e.g., words “credit” and “card”), and we aim to capture those bigrams and use them in keyword-based user profiles. To detect bigrams, we use a procedure that counts the unigram and bigram appearances, and for each combination of words wi and wj it calculates the following score: score(wi, wj ) =
count(wi, wj ) count(wi)count(wj ) .
(4.1) Finally, bigrams with a score above a certain threshold were extracted from the text in blog contents. Tags, on the other hand, naturally form n-grams, and we extracted them in their original form. 4.3
Available data sources were used to create user profiles. In particular, we extracted three distinct groups of user-related data: 1) declared; 2) content of posts; and 3) actions. The specific components included in each of the data groups are listed in Table 1. From each group we extracted features to represent the users as described below.
Declared data consists of information which a user provided during sign-up, including keywords from the blog title and blog description, where keywords were extracted using the method in Section 4.2. To create user profiles we kept the most frequent keywords from blog titles and descriptions, after removing stopwords such as “a”, “the”, “where”, “in”. We counted the keyword frequency in a user’s blog title and description, and stored the count along with the time stamp of the latest log-in as a part of the user profile.
Content features were formed from the textual contents of posts which a user either created or reblogged. The main content feature types included: 1) post tags; 2) keywords from the post title and body; 3) keywords from photo post captions; and 4) artist names from audio posts. Tags in posts were not tokenized, instead they were used in the form they appear, e.g., tag “food for a vegan” was one keyword. On the other hand, in order to extract keywords from text appearing in text post content, we again used the method from Section 4.2. We kept only the most frequently occurring keywords, excluding stopwords. In addition, we used the most popular artist names as keywords. In this way, we collected several millions of distinct keywords that were tags within a single post tj-‐n … tj-‐1 tj+1 … tj+n tags within a single post tj-‐n … tj-‐1 tj+1 … tj+n Projec(on
tj j-‐th tag Projec/on
tj c1 … ck j-‐th tag j-‐th tag categories used to obtain rich representation of user profiles. To illustrate content keyword extraction from our dataset, consider that user ui at time stamp t used tag #hp five times and tag #nba eight times, keyword football two times in post titles and posted an audio post with a song from artist Shakira ten times, then the resulting user profile would be: ui = {tag : {#hp, t : 5, #nba, t : 8}, title : {f ootball, t : 2}, artist : {shakira, t : 10}}.
Action features include follows, likes, and reblogs. If user ui follows user uj at time stamp t, we create an indicator feature f ollows : {j, t : 1} and add it to the ui’s user profile. Similarly, if user ui likes or reblogs user uj’s post, we create a feature that keeps record of the number of likes m and reblogs n, as likes : {j, t : m}, reblogs : {j, t : n}, respectively, and update the user profile accordingly.
The described approach resulted in user profiles for a total of 81.8 million users. The total number of unique features was 1.4 million, and an average user had 379.9 non-zero features. Most of the features described above are represented as either binary indicators or counts of occurrences.
The goal of our work is to infer user demographics
(described in the following section) and identify user groups
with interests in certain topics, such as music, travel,
cooking or books, in order to allow advertisers to target
segmented Tumblr audiences. As the topics may be defined at
various levels of granularity, to avoid sparsity problems while
still providing useful and actionable interest categories, user
interests are often classified into a pre-determined
hierarchical interest taxonomy that the advertisers commonly use.
However, to be able to create effective user interest
classifiers, one requires a sufficient amount of labeled data. Yet,
for the problem of the scale of Tumblr interest prediction,
this can be a daunting task for human editors. For that
reason we propose a novel semi-supervised classification
approach, based on the recently proposed word2vec model [
We decided to classify keywords into the General Interest Taxonomy (GIT), used by the Yahoo Gemini advertising platform for native advertising9. The GIT is carefully derived based on Interactive Advertising Bureau (IAB) taxonomy recommendations, in order to meet advertiser needs and protect Yahoo’s interests. The GIT has a two-level hierarchical structure, such that advertisers can adjust the audience reach by utilizing broader or narrower interest categories. The top level of the taxonomy contains 23 nodes (e.g., “Automotive”, “Business”, “Pets”, “Travel”), while the second level contains 130 nodes which represent more precise interests (e.g., “Automotive/SUV”, “Automotive/Luxury”, “Pets/Dogs”). 5.2
In this section, we present a novel classification approach
based on the recently proposed skip-gram model [
We consider the task of tag classification, where the goal is to classify tags into a pre-defined taxonomy of interest categories. In order to address this problem, we propose to learn tag representation in a low-dimensional space using neural language models that are applied to historical Tumblr posts. Let us assume that we are given N posts. In the post logs found in Firehose, every post pi is recorded along with the tags tij , j = 1...M . We collected data in the form pi = {tij , j = 1...Mi}, where Mi represents the number of N tags in the i-th post. Given the data set D = ∪i=1pi, the objective is to find a representation of tags in which semantically similar tags are nearby in the representation space. For this purpose, we extend ideas originating from recently proposed language models, as described in the remainder of this section.
The skip-gram (SG) model involves learning
representations of tags in a low-dimensional space from post logs in
an unsupervised fashion, by using the notion of a blog post
as a “sentence” and the tags within the post as “words”,
borrowing the terminology from the Natural Language
Processing (NLP) domain (see Figure 5). Tag representations
using the skip-gram model [
log P(tj+m|tj).
(5.1) kg_coks chocc where vt and vt0 are the input and output vector representations of tag t of user-specified dimensionality d, n defines the length of the context for tag sequences, and T is the number of unique tags in the vocabulary. From equation (5.2) we see that tags that co-occur often and tags with similar contexts (i.e., with similar neighboring tags) will have similar vector representations as learned by the word2vec model.
The semi-supervised skip-gram (SS-SG) model assumes that some tags are labeled with categories from the GIT taxonomy. Next, we introduce a dummy category vector for each node of the taxonomy, and leverage the tag contexts in blog posts to jointly learn tag vectors and category vectors in the same feature space. Given such a setup, after learning the representations, every tag from the vocabulary can be categorized by simply looking up the closest category vector in the joint embedding space.
Specifically, given the labeled tags, we extend D to obtain data set Dss where categories are imputed into post “sentences” pi, where available. In particular, labeled tags are accompanied by assigned categories, and every time a vector of a labeled central tag tj is updated to predict the surrounding tags, the vectors of categories assigned to tj are updated as well. More formally, assuming the central tag tj is labeled with Cj of C categories in total, ζj = {c1, . . . , cCj }, the semi-supervised skip-gram learns tag and category representations by maximizing the following objective function, X X
X p∈Dss tj∈p −n≤m≤n,m6=0 log P(tj+m|tj)+ X log P(tj+m|c) .
c∈ζj (5.3) The probability P(tj+m|c) of observing tag tj+m, given label c of the current tag tj, is defined using the soft-max, P(tj+m|c) = exp(vc>vt0j+m ) PT t=1 exp(vc>vt0) .
(5.4) This procedure allows us to seamlessly incorporate labeled and unlabeled data, and learn tag and category vectors in the joint embedding space. Then, the classification of tags amounts to a simple nearest-neighbor search among the category vectors. In Figure 6 we show the graphical representation of the semi-supervised skip-gram model. 5.2.1 Training
The models are optimized using stochastic gradient
ascent, suitable for large-scale problems. However,
computation of gradients ∇L in (5.1) and (5.3) is proportional to the
vocabulary size T , which may be computationally expensive
in practical tasks seeing as how T could easily reach several
million tags. As an alternative, we used a negative sampling
approach [
The data set used during the model training comprised 6.8 billion posts that contained tags. To collect category labels for some of the tags, we sorted the tags in a decreasing order of popularity and the editors labeled the top ones with one or more categories. Following this process, we obtained 8,400 categorized tags. We show several examples in Table 2.
The models were trained using a machine with 96GB of
RAM memory and 24 cores. The dimensionality of the
embedding space was set to d = 300, and the context
neighborhood size was set to n = 5. Finally, we used 10
negative samples in each vector update for negative sampling.
Similarly to the approach in [
When the vector representations of all tags are learned, we can find similar tags for a given tag by straightforward extracted from an activity is of class k, and 0 otherwise. In addition, we used the time stamp tact, representing the day in which the activity happened, to exponentially decay less recent activities to account for passing interest (we used α = 0.99 in our experiments). Note that the set Ai, in addition to the user’s original content, also included posts reblogged by user ui.
t
The value of ui,cat=k represents an exponentially timedecayed count of all the activities in the k-th category. To effectively store the user profile for interest targeting, instead of storing all possible activities and their time stamps, we maintain a decayed sum of the activities and update the t ui,cat=k daily.
Using this approach, we are able to qualify top K users in t each category by sorting the interest score ui,cat=k. Depending on the advertiser’s goals and the category, the choice of K varies from campaign to campaign. We note that a single user can be qualified into one or more interest categories (e.g., a user can be categorized in “Sport”, “Sport/Basketball”, and “Health and Nutrition/Vitamins” at the same time), and, when the system was deployed, each user was assigned to 13 categories on average. An example of user profiles qualified into certain categories is given in Table 5. 5.3.1
Leveraging the follower graph
To be able to target Tumblr users who do not create much content, but who actively follow and engage with other blogs, we leverage the follower graph to create additional categorized features. Using equation 5.5 we can identify frequent bloggers in certain interest categories by focusing on a small percentage of users with a maximum ui,cat=k. Following and liking posts created by social influencers in the k-th category serves as additional evidence of one’s interest in that category.
In each interest category, we label the 5% of users with the highest number of activities in that category as frequent bloggers. Next, we update the interest score of all users ui in the k-th category, in the following manner
t ui,cat=k+ =
X X α(t−teng) weng I(b is of class k), (5.6) where Fi is a set of all frequent blogs followed by user ui, eng are all engagements with the b-th blog, i.e. likes or follow actions, along with their weights weng (e.g., if the b∈Fi eng∈b posts created by the b-th blog were liked ten times, then the value was set to 10; if a user followed the b-th blog, then the value was set to 1), while the indicator function I(·) returns 1 if the blog is of class k, and 0 otherwise. Similarly to other activities, we applied the exponential decay to the sum, based on the time stamps of follow and like actions.
We have observed that additional signals, in the form of follow and like engagement with frequent bloggers, increase our segment sizes, making it possible to efficiently target a greater number of users. 5.4
In order to evaluate the generated user interest segments, we performed online A/B testing and worked with several advertisers who ran concurrent interest-targeted and untargeted campaigns. We tracked user engagement with their ads in terms of sponsored post likes, reblogs and follows, and show the results for 8 targeting campaigns in Table 6. We observed an average increase of 20% in user engagement (aggregate of 3 metrics) with sponsored posts in comparison to untargeted campaigns. This performance result represents a significant improvement over the baseline approach. 6.
In this section, we explain the details of our gender prediction model, based on the user profiles described in the previous sections. We first describe the generation process of a golden set of labeled users, which is used to train a predictive model that generalizes well on the remaining unlabeled users. This is followed by the model’s description and a discussion of the results. #Microposts2016 6th Workshop on Making Sense of Microposts
In order to train the machine learning method for gender prediction, in addition to user profiles, we also require labels that present the ground truth (i.e., “male” or “female”). However, Tumblr does not collect gender information when users sign-up, leaving open the question of how to obtain such data.
To address this problem, we proposed to leverage highly informative blog description data in order to infer user gender information. In particular, very often users declare their name in the blog description, as illustrated in Figure 1. To extract the users’ declared names, we used several regular expression rules that we found to result in very high precision. The obtained results from a large set of name-matching regular expressions were editorially tested for quality. It was found that regular expressions reported in Table 7 yielded the most reliable extracted names (valid names were extracted in more than 95% of the cases). Next, in order to generate the gender ground truth, we used the US census data of popular baby names10 from year 1880 to 2013 to create a “name → gender” mapping. Seeing as how certain names are given to both males and females, we used the empirical counts of babies with certain names from census data to generate the labels. More specifically, we used male/female empirical ratios as soft labels, with 1 indicating 100% confidence in a male and 0 indicating 100% confidence in a female name. 6.2
Let Dg = {(xi, yi), i = 1, ..., N } denote our gender data
set, where N is the total number of labeled users, xi is the
K-dimensional user feature vector generated from the user
profiles, and yi ∈ [
Our goal is to learn a gender-predictive model, f : x → y. As a classification model, we used logistic regression, parameterized by weight vector w. We assume that the posterior gender probabilities can be estimated as a linear 10www.ssa.gov/oact/babynames/limits.html and P(y = 0|x) = 1 − P(y = 1|x). To estimate the parameters w, we minimize the following loss function,
N min 1 X w∈RK N i=1
yi − f (xi, w) 2 + λkwk1, where hyper-parameter λ controls the `1-regularization, introduced to induce sparsity in the parameter vector and reduce the feature space to a subset of features that are the most predictive. For data sets with a large number of features, as we have in our use case, it is common that many features are not useful for producing the desired learning result. For this reason, the `1-regularization was a critical part of our training procedure. In addition, we experimentally observed that the model generalizes better when we trained an initial model with `1-regularization to find which features have non-zero weight, and then do another round of training without `1-regularization, by only using features with non-zero weights from the first round to learn a better classifier.
Given a trained logistic regression model, the posterior
class probabilities are estimated as f (xi, w) ∈ [
To evaluate the accuracy of our gender prediction
framework, we trained a logistic regression model on 70% of the
golden set and tested on the remaining 30%. We used the
Vowpal Wabbit [
In addition to evaluation on the hold-out set, we also editorially evaluated gender predictions on the unlabeled data set of user profiles. We randomly picked 1,007 gender predictions from the population of 64.1 million users and asked editors to visit their profiles and verify their gender. They were instructed to mark our predictions as “correct”, “incorrect”, or “not sure”. The “not sure” grade is to be used when the visual inspection of a profile is inconclusive, as we found was often the case. The editorial judgment came back with 573 “correct” (429 females and 144 males), 9 “incorrect”, and 425 “not sure” grades (see Table 9). The fact that there are so many “not sure” grades indicates that in many cases it is hard to infer the gender even after manual efforts, further indicating the benefits of the proposed approach and its superior performance in comparison to humans. Finally, we retrained the model with 100% of the golden set and deployed it in Yahoo production systems.
A demonstration video of the most predictive tags in each gender group is available online at https://www. youtube.com/watch?v=jXGJ0TpOlhg.
Due to the rapid growth of Tumblr and the large number of activities generated by the existing users, we implemented daily scoring of users in Yahoo production servers. We store the activities, i.e. raw counts as well as decayed counts, in Hive tables11 for efficient retrieval. The decayed counts used in interest prediction are updated on a daily basis by multiplying the old feature values by the decay factor α and adding new activities. In order to infer the gender of new users we implemented daily scoring by leveraging MapReduce on Hadoop12. Both interest and gender models are retrained on a regular basis.
After thorough editorial evaluation of the inferred gender and interest targeting, both targeting frameworks were enabled through Gemini self-serve tool13. Advertisers can choose to use gender and/or interest targeting with custom segment sizes, allowing for effective targeting campaigns.
We have presented the steps in the development of a largescale Tumblr gender and interest targeting framework, where we used historical Tumblr activities to create rich user profiles. We described the methodology, including a novel semisupervised neural language model, as well as the high-level implementation details behind the deployed system. Currently, our gender and interest predictions cover users that generate more than 90% of Tumblr’s daily activities, and are heavily leveraged by advertisers. In our ongoing work, we are concentrating on creating custom keyword-targeted advertising segments, specifically tailored for a particular advertiser, which include addressing of the problems of keyword discovery and expansion.