Supporting Serendipitous Discovery and Balanced Analysis of Online Product Reviews with Interaction-Driven Metrics and Bias-Mitigating Suggestions

Title of the Paper

Supporting Serendipitous Discovery and Balanced Analysis of Online Product Reviews with Interaction-Driven Metrics and Bias-Mitigating Suggestions

Paper Information

• Subject Area: Human-Computer Interaction and Text Analytics, with a focus on exploration and decision support for online product reviews.
• Keywords: serendipity, product review exploration, bias mitigation model, interaction-driven metrics, text analytics system, data visualization, decision-making, review analysis, user trust, user agency.

Research Background and Problem

• Identified Problems or Challenges:

  • The vast quantity of online product reviews makes it difficult for most users to comprehensively explore and analyze them.
  • Users are prone to cognitive biases, often seeking reviews that confirm their assumptions while ignoring other information.
  • Existing technologies primarily focus on numerical or categorical data, which are not well-suited for text-heavy product reviews.

• Significance:

  • Online reviews play a crucial role in purchase decisions. Enhancing their explorability can help consumers make more comprehensive, data-driven decisions.
  • Bias in review exploration may lead to users misjudging a product's value, affecting their purchasing experience.

• Research Motivation and Related Work:

  • Expanding the breadth of coverage and information exploration is a key focus in the fields of text analytics and information visualization.
  • Previous studies have primarily concentrated on navigating aggregated data in visual analysis, while exploratory analysis at the text level remains underdeveloped.
  • This study integrates data visualization and hyper-precision goals from recommendation systems to extend methods for "discovering serendipitous information."

Proposed Solution

• Proposed Method or Solution:

  • Development of an interactive text analytics system, Serendyze, designed to support users in exploring and analyzing online product reviews.
  • Introduction of two core intervention mechanisms:
    1. Exploration Metrics: Three metrics to help users track exploration patterns—Visit, Coverage, and Distribution.
    2. Bias Mitigation Model: Suggestions that present reviews with semantic and emotional differences from those already read by the user, helping them gain broader knowledge.

• Innovations:

  • Combining "serendipitous discovery" with text analytics to provide users with multi-perspective review content, avoiding exploration bias.
  • Introducing efficient guidance mechanisms at the text level, distinct from traditional numerical data exploration methods.

• Implementation Steps and Techniques:

  • Exploration Metrics:

    • Define the Visit metric based on direct reviews read by the user.
    • Use Doc2Vec embeddings and semantic similarity measures to calculate the Coverage metric.
    • Set the Distribution metric based on the difference between the user's exploration of various sentiments and the actual data distribution.

  • Bias Mitigation Model:

    • Generate diverse recommendations using document semantic embeddings and cosine similarity to avoid biased reading.
    • Optimize algorithmic suggestions based on sentiment distribution to ensure coverage of diverse viewpoints.

Research Outcomes

• Specific Results:

  • Enhanced Exploration Coverage: Users employing exploration metrics covered more reviews in less time and explored in a more balanced manner.
  • Increased User Confidence: Bias-mitigating suggestions improved users' confidence in the comprehensiveness of their reading and enhanced the reliability of their final decisions.
  • Practical System Utility: Compared to the control group, users of Serendyze significantly improved their review coverage and depth of knowledge.

• Advantages Compared to Existing Solutions:

  • Overcomes the limitations of existing research that overly focuses on aggregate statistics in text analysis, enabling users to gain deeper insights directly from review content.
  • Reduces bias issues often caused by traditional recommendation algorithms through bias-mitigating suggestions.

• Experimental or Evaluation Results:

  • In an experiment with 100 participants, the group using exploration metrics and the bias mitigation model covered an average of 234 reviews, compared to 66 reviews in the baseline group.
  • Participants exhibited a more balanced distribution when reading positive, negative, and neutral reviews, avoiding biased reading patterns.
  • User feedback indicated that exploration metrics enhanced decision comprehensiveness, while the suggestion feature supplemented unknown information.

• Limitations and Future Directions:

  • The system may experience delays under high computational loads, necessitating optimization of computational processes and server responsiveness.
  • Some users perceived the suggestions as too brief or irrelevant, leading to mistrust; future work should improve suggestion quality evaluation mechanisms.
  • Before real-world deployment, the system needs expanded functionality to include product visualizations and cross-domain experiments, such as public opinion analysis or social media content exploration.