UC San Diego

The development of personalized systems, driven by sophisticated machine learning models, has notably enriched user experiences across various digital interfaces. However, these systems often obscure the rationale behind personalized recommendations, creating a pressing need for enhanced explainability. We present a comprehensive framework aimed at bridging this explainability gap by systematically extracting, understanding, and demonstrating key information to users.

First, the framework starts with the extraction of information using named entity tagging. This step facilitates the identification and extraction of significant entities and terms from vast datasets. The precision in extraction is crucial as it directly impacts the quality of understanding and explanation in subsequent phases. Upon successful extraction, the framework transitions to the understanding phase, where the unsupervised contrastive learning model, UCTopic, is employed. This model analyzes the extracted phrases, diving deep into their semantic and thematic contexts. It generates context-aware phrase representations and mines topics, thereby elucidating the thematic essence and semantic correlations encapsulated within the data. Finally, we leverage the strength of various models to generate coherent and intuitive explanations. These generative models can be categorized based on topic, keyphrase, or multi-modality. The generated explanations provide a clear rationale behind the recommendations, making them easily interpretable and relatable to the users.

In summation, our research improves the level of transparency and interpretability inherent in personalized systems. The empirical assessments show the effectiveness of our research in supporting explainability, thereby having a more transparent and user-aligned experience. Through this endeavor, our research substantially improves explainability in personalized systems forward, resulting in a more intuitive and user-friendly interaction paradigm.