How Pinterest Fights Spam Using Machine Learning
Vishwakarma Singh | Trust and Safety Machine Learning Lead
Hundreds of millions of people regularly visit Pinterest to visually discover inspiring ideas among billions of Pins. Inspiration is a high bar and we must be vigilant in ensuring that Pinners don’t see spam, harmful content or misinformation. To enforce our community policies and maintain an inspiring environment, we use the latest in machine learning technology to build automated systems that swiftly detect and act against both spammy content and spammers.
Our anti-spam system consists of both reactive and proactive components to effectively counter adversarial abusers — users who intentionally try to evade the system. Our proactive system consists of sophisticated machine learning models, whereas the reactive system includes both rules executed in a real-time rules engine and lightweight machine learning models. We not only use the latest modeling techniques but also iterate on these models at regular intervals by adding new data and exploring new technical breakthroughs to either maintain or improve their performance over time to effectively address spam.
One tactic malicious actors enact is misusing a Pin’s image and linking to a malicious external website. Our models detect spam vectors, like Pin links, as well as users engaging in spammy behaviors. We quickly limit distribution of Pins with spam links and take direct action against users identified with a high confidence to be engaging in spammy behavior. We perform a manual review for those identified with low confidence to limit false positives, and we notify users of our actions to maintain transparency and also provide an option of appeal against our decision.
Machine Learning Models
Spam Domain Model
We proactively identify spam Pin links using a Deep Neural Network classifier (shown in Figure 1). To maximize impact, our model learns to classify a domain as spam rather than a link. We apply the same enforcement to all Pins with links belonging to the same domain. This model is trained interactively on manually labeled domains to achieve a higher recall and lower false positive rate. We use features created from links, web page text and media, user-domain interactions, and user behavior as inputs. For each domain, we sample links and webpages to create features. We semantically split links into semantic tokens and use only frequent tokens as features. We analyze outlying patterns in user actions over time to create behavioral features. This model is periodically batch inferred at scale by a PySpark job using Tensorflow, Spark SQL, and a UDF.
Spam User Model
Identifying users engaging in spam activities is the ultimate solution for fighting spam, but it is extremely hard to achieve. We leverage both supervised and unsupervised models to build an effective spam user identification system.
Our spam user classification model is a Deep Neural Network (shown in Figure 2) and is part of our proactive system. It is trained using synthetically labeled data generated with minimal human supervision to ensure quality. We use features created from user attributes and their past behaviors as inputs. We also use user-domain interaction, summarized as a domain scores distribution for each user where domain scores are reused from the spam domain model, as an input. This model is periodically batch inferred to score millions of Pinners by a PySpark job using Tensorflow, Spark SQL, and a UDF.
We have developed lightweight clustering models for early detection of suspicious users and bots. This technique also addresses gaps in our classification models, which are unaware of emerging patterns unless re-trained with fresh labeled data. We cluster users on attributes which can successfully isolate suspicious groups with high accuracy. Experts identify these attributes by exploring the behavior of suspicious users and their use of resources for creating spammy content. This model is implemented using PySpark and SparkSQL and executes daily.
Spam User-Domain Model
Interactions of users with domains are explicitly captured by a heterogeneous bipartite graph as shown in Figure 3. We represent users and domains as nodes in the graph and create an edge between a user and a domain if the user has created or saved a Pin with the domain’s link. This graph facilitates simultaneous identification of spam users and domains using a semi-supervised learning. We use a small set of labeled users and domains to run a label propagation algorithm and learn scores for the unlabeled users and domains. We implement this iterative algorithm in Spark and run it periodically.
We measure spam prevalence on Pinterest by computing the number of Pin impressions which either have spam links or have been created by users engaging in spammy activities. We periodically sample and manually review both impressed Pins and users. We scaled our measurement by starting to sample and review from highly impressed head domains and then extended the coverage to tail domains over a period of time. These samples are used for measuring overall spam prevalence as well as training our machine learning models.
Pinterest’s mission is to bring everyone the inspiration to create a life they love. We strive to protect our Pinners’ experiences by swiftly and appropriately acting against malicious users and spam content as identified by our array of latest machine learning models. We plan to keep investing in evolving our community guidelines and technology to address inevitably emerging challenges and bring the best experience to our millions of valued users.
Thanks to Yuanfang Song, Omkar Panhalkar, Rundong Liu, Qinglong Zeng, Attila Dobi, Abhijit Mahabal, Alok Singhal, Maisy Samuelson, and the rest of the Trust and Safety team for their contributions in developing machine learning models for spam! Thanks to Harry Shamansky for helping with the publication of the blog post!