Detecting Spam Game Reviews on Steam with a Semi-Supervised Approach

The proliferation of deceptive content on digital storefronts undermines consumer trust and distorts market signals. To address this, a semi-supervised framework utilizing Positive-Unlabelled learning provides a scalable and accurate mechanism for identifying spam game reviews on the Steam platform. By categorizing problematic content into non-reviews, duplicates, and untruthful entries, this approach effectively mitigates the limitations inherent in manual annotation, which is frequently both time-consuming and prone to human error.

The methodology relies on a comprehensive feature extraction process that incorporates stylistic, syntactic, sentiment-based, and platform-specific raw data, alongside Doc2vec embeddings. Applied to a dataset of 33,450 Steam reviews, the model successfully identified approximately 15% of the content as spam, a figure that aligns closely with established industry benchmarks. Experimental evaluations against a Gold-Standard dataset confirm the efficacy of this approach, with the PU-bagging method achieving an F1 score of up to 80%. This performance demonstrates that the model is not only robust when working with limited positive training samples but also consistently outperforms human evaluators in detection accuracy.

Prioritizing high recall is essential for maintaining the integrity of review ecosystems, as the negative impact of deceptive content necessitates comprehensive filtering. While the current model offers a significant advancement in automated moderation, future refinements will focus on improving the interpretation of non-traditional review elements, such as emojis and numerical ratings. By automating the detection process, this framework offers a viable path forward for platforms seeking to maintain authentic user feedback loops without the prohibitive costs associated with manual oversight.