Expanded Tag Genomes for Cross-Domain Recommendation

This repository is designed to tackle the item-tag prediction problem for books and movies. An item refers to either a book or a movie, and the goal is to predict the relevance of tags for items using multiple feature sets and models.

The repository consists of two main parts:

• Evaluation: Selects items for evaluation, prepares data, generates features and evaluates multiple models.
• Score Generation: Uses training data (from Evaluation or downloaded directly) to generate item-tag scores.

Dataset

The dataset generated in this project can be accessed here.

License

At this stage, the code and dataset in this repository are provided for viewing purposes only.
You may not use, copy, modify, or distribute any part of the code or dataset at this time.

The materials are temporarily restricted while the paper associated with this repository is under review.
Once the paper is accepted for publication, both the code and the dataset will be released under a Creative Commons license that allows legal and free use in academic and research settings.

---

Feature Sets

The project uses three feature sets: original, core and original_core.

Original Feature Set

This feature set combines information about tag applications, user ratings, and user reviews to predict tag relevance for items.

• Tag applications

  • tag_exists: Indicates whether a tag has been applied to an item, with 1 meaning applied and 0 meaning not applied
  • lsi_tags_75: Measures similarity between a tag and an item using latent semantic indexing (LSI), where items are represented by their applied tags

• User ratings

  • rating_similarity: Cosine similarity between the item's ratings and the aggregated ratings of items linked to the tag
  • avg_rating: The average rating given by users to the item.

• User reviews

  • log_IMDB: Log-scaled frequency of the tag appearing in user reviews for the item, calculated after applying stemming to both tags and reviews.
  • log_IMDB_nostem: The same as above, but calculated without applying stemming, so original word forms are preserved.
  • lsi_imdb_175: LSI-based similarity between a tag and an item, where items are represented using the bag-of-words from their reviews

• Feature interactions

  • tag_prob: An estimated relevance score for the relationship between a tag and an item. It is computed using logistic regression, where the target variable is tag_exists and all other features are used as input.

Core Feature Set

The core feature set focuses on tag applications, user ratings, and user reviews using alternative metrics:

• User ratings
  • avg_rating: Mean rating of the item
  • pop: Item popularity, calculated as log of number of ratings
• User reviews
  • lemma_review_mentions: Log-scaled frequency of the tag in lemmatized reviews
  • raw_review_mentions: Same as above, but without lemmatization
  • lemma_max_tfidf, lemma_mean_tfidf: TF-IDF scores per lemmatized review, aggregated by max or mean
  • raw_max_tfidf, raw_mean_tfidf: Same as above, but without lemmatization
  • bert_avg_sim, bert_max_sim: Cosine similarity between BERT embeddings of tag and item reviews, aggregated by average or max
  • bert_description, bert_highlights: Cosine similarity between BERT embeddings of tag and item metadata (description and highlights)

Original_Core Feature Set

Represents the combination of core and original features