TL;DR
This paper introduces a novel, "off-the-shelf" method for authorship attribution that treats it as an anomaly detection problem. By combining semi-supervised anomaly detection with multiple kernel learning (MKL), their approach automatically learns the optimal combination of features (kernels), eliminating the need for manual feature selection and making it broadly applicable.
What Problem Does This Paper Address?
Authorship attribution, the task of identifying the author of a given text, is a long-standing problem with applications ranging from plagiarism detection to historical text analysis. A major hurdle in this field is the "curse of dimensionality" and the challenge of selecting the right set of features that effectively capture an author's unique writing style. Existing methods often require extensive manual feature engineering, making them difficult to adapt to new authors or tasks. Furthermore, the choice between in-sample (training and test data from the same distribution) and out-of-sample (test data from a different distribution) settings significantly impacts performance, and traditional methods struggle to bridge this gap.
Methods
The authors propose a unified framework that casts authorship attribution as an anomaly detection problem. For each author, they learn a model that defines a "region" in feature space containing that author's writings, treating other texts as outliers. This is achieved using a semi-supervised anomaly detection approach, which can leverage both labeled (known author) and unlabeled (disputed author) documents.
The core innovation lies in the integration of Multiple Kernel Learning (MKL). Instead of manually selecting features (which are represented as kernels), MKL allows the model to learn an optimal linear combination of multiple kernels. This means the algorithm automatically determines the best way to weight different feature sets (e.g., function words, part-of-speech tags, n-grams) to best distinguish authors. This approach addresses the problem of unknown optimal feature representations and makes the method generally deployable without requiring expert knowledge for feature selection.
Datasets
- Reuters-50-50 Corpus: A subset of this corpus was used, containing 1000 articles from 10 authors. This dataset was split into training (90%) and testing (10%) sets, with a 10-fold cross-validation performed on the training set for model selection. The experiments used a transductive setting where training instances were labeled, and holdout/test articles were unlabeled.
- Federalist Papers: This dataset consists of 85 articles. The authors used 73 articles with known authorship for training and the 12 disputed papers for testing. They assumed that all 12 disputed papers were written by Madison, based on prior studies. Experiments were repeated five times with randomly drawn training and holdout sets.
Key Insights and Ideas
- Authorship Attribution as Anomaly Detection: Framing the problem as finding an "author's region" in feature space simplifies the task and allows for a more robust approach, especially with limited data.
- The Power of Multiple Kernel Learning (MKL): MKL is crucial for making authorship attribution "off-the-shelf." By automatically learning the optimal combination of diverse feature sets (kernels), it removes the burden of manual feature engineering and selection, which is a major bottleneck in traditional methods.
- Semi-Supervised Learning Bridges the Gap: The integration of semi-supervised learning allows the model to effectively utilize both labeled and unlabeled data, improving performance in both in-sample and out-of-sample scenarios.
- Author-Specific Feature Combinations: The learned kernel mixture coefficients (shown in Figure 2) reveal that different authors benefit from different combinations of features, highlighting the adaptive nature of the MKL approach.
Key Results
- Superior Performance: The proposed MKL-based approach consistently outperformed baseline single-kernel methods (like SSAD) on the Reuters-50-50 dataset, achieving higher micro- and macro-F1 scores across various kernel combinations and p-norms.
- Confirmation of Prior Findings: On the Federalist Papers, the MKL method confirmed the prevailing assumption that the disputed papers were written by Madison, aligning with the results of some baseline methods but offering a more principled and automated way to arrive at this conclusion.
- Robustness: The method demonstrated robustness across different datasets and experimental settings. The authors note that using MKL with a sum of kernels (as a baseline) performed worse than using a single, well-chosen kernel, underscoring the importance of learning the optimal mixture rather than just combining features.
Limitations and Future Work
- Computational Cost: While the method is "off-the-shelf" in terms of feature selection, the optimization process for MKL can be computationally intensive, especially with a large number of kernels.
- Interpretability of MKL Weights: While the learned weights (βt) indicate the importance of each kernel, interpreting the exact contribution of individual features within a kernel might still require further analysis.
- Future Work: The authors suggest exploring more general p-norms for regularization of mixing coefficients and further investigating the interpretability of the learned kernel mixtures. The general framework could also be extended to other text classification tasks.
Summary updated