To be effective in detecting spam in online content sharing networks, it is necessary that techniques used to detect spam have good precision, high recall, and the ability to adapt to new types of spam. A bottleneck in developing such machine learning techniques is the lack of availability of high quality labeled training data. Human labeling to obtain high quality labeled data is expensive and not scalable. Current approaches such as unsupervised learning or semi-supervised learning can only produce low quality labels.

Generally, the present disclosure is directed to a weak supervision approach to train a machine learning model to detect spam content items. Weak labels are generated for content items in training data using various techniques such as rules that encode domain knowledge and/or anomaly detection techniques such as unsupervised machine learning/ clustering or semi-supervised machine learning. The accuracy of the various techniques is estimated based on observed agreements/ disagreements in the weak labels. The weak labels are combined into a single value (e.g., per content item) that is used as a probabilistic training label to train a machine learning model using supervised learning that is noise aware. In the training, a penalty is applied for deviation from the probabilistic label such that the penalty is higher for a label associated with a higher confidence and lower for a label associated with a lower confidence. The model thus trained can be used to detect spam content.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.