Experimental analysis on cross domain preferences association and rating prediction

Zhenhua Dong, Qian Zhao2012 · DOI

TL;DR

This paper explores how user preferences for tags in one domain (like movies) can predict preferences in another (like books or music), and demonstrates the feasibility of cross-domain rating prediction using a KNN-based approach. The findings suggest that tag associations are a strong indicator of cross-domain user behavior, paving the way for more comprehensive recommendation systems.

What Problem Does This Paper Address?

In today's digital world, users interact with a vast array of online content across different domains – movies, books, music, and more. Recommender systems are crucial for helping users navigate this information overload, but most systems operate within a single domain. This creates a challenge: how can we leverage a user's known preferences in one domain to make relevant recommendations in another? This "cross-domain recommendation" problem is significant because it can help address the "cold-start" issue (recommending to new users or for new items) and enhance the novelty and serendipity of recommendations. The authors aim to understand the underlying connections between user preferences across domains and to build a practical system for cross-domain rating prediction.

Methods

The paper tackles two main problems:

  1. Cross-Domain Preference Association: They use "folksonomy" (user-generated tags) to uncover relationships between user preferences across different domains.
  2. Tag Preference Calculation: First, they calculate the relevance of a tag to an item using TF-IDF. Then, they compute a user's preference for a tag based on their ratings of items associated with that tag.
  3. Association Rule Mining: Using these user tag preferences, they mine association rules (e.g., "if a user likes tag X in domain A, they are likely to like tag Y in domain B"). They use standard metrics like support, confidence, and lift to identify interesting rules.
  4. Cross-Domain Rating Prediction: They employ a user-based K-Nearest Neighbors (KNN) model for this task.
  5. Neighbor Computation: Given a "candidate" domain (where user preferences are known) and a "target" domain (where predictions are needed), they first build a user-item rating matrix for the candidate domain. They then compute the similarity between users using adjusted cosine similarity and identify neighbors (users with similar tastes).
  6. Prediction: Once neighbors are identified, they use their ratings in the target domain to predict a user's rating for an item in that target domain. The formula essentially averages the weighted ratings of neighbors, where weights are based on similarity.

Datasets

The experiments were conducted on data crawled from Douban, a Chinese social website focusing on cultural products. They collected rating and tag information for movies, books, and music.

  • Dataset Size: Approximately 2,000 users and 49,000 items across the three domains.
  • Data Characteristics: The paper provides details on item quantities, rating quantities, average ratings, distinct tag quantities, and tag applications for each domain.

Key Insights and Ideas

  • Tags as Bridges: User-generated tags are powerful indicators of preferences that transcend domain boundaries. The correlation in tag usage between domains suggests a shared underlying user interest.
  • Movie Domain as a Strong Predictor: Interestingly, the movie domain appears to be a particularly effective candidate domain for generating neighbors, leading to lower predictive errors even when predicting ratings in book or music domains. This suggests that movie-watching habits might be a good proxy for broader taste.
  • Cross-Domain vs. Single-Domain Performance: While cross-domain prediction is viable, the paper notes that it doesn't always outperform single-domain prediction. The effectiveness is highly dependent on the choice of candidate and target domains.
  • Limitations of Simple Similarity: The authors acknowledge that simple item-based similarity can be coarse-grained. Users might rate an item similarly but for different reasons, leading to prediction deviations.

Key Results

  • Preference Association: The paper successfully mined numerous interesting tag preference association rules across book-movie, movie-music, and book-music domains. For example, they found that users who like "Japanese Comics" in books are highly likely to like "Japanese Animation" in movies (70.15% correlation).
  • Rating Prediction Accuracy:
    • The best cross-domain prediction results (lowest MAE and RMSE) were often achieved when using the movie domain as the candidate domain, even for predicting book and music ratings. For instance, predicting book ratings with movie domain neighbors yielded an MAE of 0.8128 and RMSE of 1.0475.
  • Predicting movie ratings using movie domain neighbors as the candidate resulted in the lowest MAE (0.7484) and RMSE (0.9523) among single-domain predictions.
  • The "Combine" condition (using all domains for both candidate and target) showed higher predictive errors (MAE=0.7735, RMSE=0.9946) compared to some of the more effective single-domain or specific cross-domain scenarios.

Limitations and Future Work

  • Understanding "Why": The paper raises the question of why these tag associations exist across domains and suggests using Topic Models to analyze tag distributions in a topic space to uncover semantic similarities.
  • Improving Prediction Accuracy: The authors identify two key areas for improvement:
  • Richer User Preferences: Moving beyond simple ratings to more sophisticated preference representations (e.g., using factor models).
  • Bridging Semantic Spaces: Addressing the deviation between semantic spaces of different domains. They propose using Topic Models to create a unified semantic space and improve neighborhood models.
  • Topic Model Limitations: They note that simply merging items for Topic Modeling might not fully capture cross-domain item heterogeneity, suggesting a need for adjusted Topic Models.

Summary updated

Abstract

Cross domain recommendation and preferences association are emerging research topics. In this paper, we will study the two topics through experimental analysis methods: firstly, we use folksonamy to analyze the preferences association among different domains; secondly, we analyze the feasibility of cross domain rating prediction based on KNN model. The experimental results report the associative tag pairs of users' preferences on items across domains. In addition, we report the cross domain prediction results here.
BibTeX
@misc{dong2012experimental,
  author = {Zhenhua Dong and Qian Zhao},
  title = {{Experimental analysis on cross domain preferences association and rating prediction}},
  year = {2012},
  doi = {10.1145/2351333.2351337},
}
Figure