A multi-view representation technique based on principal component analysis for enhanced short text clustering.

Area of Science:

• Data Mining
• Information Retrieval
• Natural Language Processing

Background:

• Clustering texts is crucial for data mining and information retrieval, aiming to group unlabeled data into meaningful sets.
• Clustering short texts (STC) presents unique challenges due to sparsity, ambiguity, and noise in the data.
• Existing STC methods often rely on single-view text representations, which are insufficient for capturing diverse textual aspects.

Purpose of the Study:

• To enhance short text clustering (STC) by proposing and evaluating a multi-view representation (MVR) approach.
• To identify the optimal combination of single-view representations for effective MVR in STC.
• To investigate the impact of different MVR strategies on the quality of text clusters.

Main Methods:

• Developed a multi-view representation (MVR) strategy by combining various single-view text representations.
• Utilized Principal Component Analysis (PCA) for fixed-length concatenation of single-view representations to create MVRs.
• Evaluated different MVR combinations on three standard datasets: Twitter, Google News, and StackOverflow.

Main Results:

• Experimental results demonstrated that multi-view representation (MVR) significantly improves short text clustering (STC) performance compared to single-view methods.
• The most effective MVR for STC was a 5-view combination, integrating BERT, GPT, TF-IDF, FastText, and GloVe representations.
• The study highlights the necessity of carefully selecting single-view representations for optimal MVR design.

Conclusions:

• Multi-view representation (MVR) is a superior approach for enhancing short text clustering (STC).
• The effectiveness of MVR hinges on the judicious selection and combination of diverse single-view text representations.
• Future research should focus on advanced methods for constructing optimal MVRs for STC.