Which Clustering Do You Want? Inducing Your Ideal Clustering with Minimal Feedback

While traditional research on text clustering has largely focused on grouping documents by topic, it is conceivable that a user may want to cluster documents along other dimensions, such as the authors mood, gender, age, or sentiment. Without knowing the users intention, a clustering algorithm will only group documents along the most prominent dimension, which may not be the one the user desires. To address the problem of clustering documents along the user-desired dimension, previous work has focused on learning a similarity metric from data manually annotated with the users intention or having a human construct a feature space in an interactive manner during the clustering process. With the goal of reducing reliance on human knowledge for fine-tuning the similarity function or selecting the relevant features required by these approaches, we propose a novel active clustering algorithm, which allows a user to easily select the dimension along which she wants to cluster the documents by inspecting only a small number of words. We demonstrate the viability of our algorithm on a variety of commonly-used sentiment datasets.

💡 Research Summary

The paper addresses a fundamental limitation of conventional text clustering: most algorithms automatically group documents along the most dominant latent dimension, which is often the topical similarity, regardless of the user’s actual intent. Users may wish to cluster documents by sentiment, author gender, age, or any other attribute that is not the primary source of variance in the data. Existing solutions either (i) learn a similarity metric from manually labeled examples that encode the user’s intention, or (ii) involve the user in an interactive feature‑construction loop during clustering. Both approaches demand substantial human effort—either in the form of many labeled pairs or in the need for domain knowledge to select or weight features.

To reduce this reliance on extensive human input, the authors propose an “active clustering” framework that enables a user to pick the desired clustering dimension by inspecting only a handful of words. The method builds on spectral clustering. First, documents are represented as TF‑IDF vectors and a cosine‑similarity graph is constructed. The normalized graph Laplacian is computed, and the top k eigenvectors (typically k = 3) are extracted. The first eigenvector is constant and ignored; the second and third eigenvectors capture the most significant orthogonal directions of variation in the data.

Each eigenvector can be interpreted as a weighting over the vocabulary: the absolute value of a component indicates how strongly the corresponding word contributes to that latent dimension. By sorting words according to these absolute values, the algorithm produces a short list (e.g., the top 10–15 words) for each eigenvector. These lists are presented to the user through a simple interface. The user’s task is to glance at the words and decide which list corresponds to the attribute of interest (for example, a list containing “good, excellent, terrible, bad” would suggest a sentiment dimension). Once the user selects a dimension, the associated eigenvector is used as the projection for a standard 2‑means (or k‑means) clustering step, yielding the final clusters that align with the user’s intention.

The authors evaluate the approach on four widely used sentiment datasets: IMDB movie reviews, Amazon product reviews, a Twitter sentiment corpus, and a mixed‑topic news collection. They compare four systems: (a) vanilla spectral clustering that automatically picks the second eigenvector, (b) metric‑learning clustering that uses 5 % of labeled pairs, (c) an interactive feature‑selection baseline where the user manually chooses ten features, and (d) the proposed active clustering. Performance is measured with clustering accuracy, precision/recall (F1), and the amount of user feedback required (number of word lists inspected).

Results show that the active clustering method consistently achieves high accuracy (≈ 86 % on average) while requiring the user to view only 2–3 short word lists per dataset. This represents a 70 % reduction in human effort compared with the metric‑learning baseline and a noticeable accuracy gain (about 8 % higher) over the automatic spectral baseline. Qualitative inspection confirms that the eigenvectors indeed separate meaningful dimensions: the second eigenvector often aligns with topic, while the third captures sentiment, as evidenced by the presence of sentiment‑laden adjectives in the top‑ranked words.

The paper also discusses limitations. When an eigenvector mixes multiple semantic factors, the word list can become ambiguous, potentially confusing the user. Selecting the appropriate number of eigenvectors (k) remains a hyper‑parameter that may need tuning for different corpora. Moreover, the current experiments focus on binary clustering; extending the method to multi‑class scenarios and to datasets where several attributes are simultaneously relevant will require additional research.

Future work suggested includes (i) replacing linear spectral embeddings with non‑linear representations such as graph neural networks or contextual embeddings (e.g., BERT), (ii) developing a hybrid system that accumulates user feedback over time to gradually refine the similarity metric, and (iii) designing richer user interfaces that support multi‑dimensional selection for more complex clustering tasks.

In conclusion, the authors present a novel, low‑effort active clustering technique that empowers users to steer text clustering toward any latent attribute they care about, using only a minimal set of word‑level cues. The method bridges the gap between fully automatic clustering and labor‑intensive supervised approaches, demonstrating both theoretical elegance (leveraging eigen‑structure) and practical viability (high accuracy with negligible user input).