Home

An in-depth commentary is provided on the Commentary page. Overall, I would suggest to match store descriptions with:

• either Term Frequency * Inverse Document Frequency (Tf-Idf),
• or a weighted average of GloVe word embeddings, with Tf-Idf reweighting, after removing some components:
  • either only sentence components,
  • or both sentence and word components (for slighly better results, by a tiny margin).

When using average of word embeddings as sentence embeddings:

• removing only sentence components provided a very large increase of the score (+105%),
• removing only word components provided a large increase of the score (+51%),
• removing both components provided a very large increase of the score (+108%),
• relying on a weighted average instead of a simple average lead to better results,
• Tf-Idf reweighting lead to better results than Smooth Inverse Frequency reweighting,
• GloVe word embeddings lead to better results than Word2Vec.

A table with scores for each major experiment is available. For each game series, the score is the number of games from this serie which are found among the top 10 most similar games (excluding the query). The higher the score, the better the retrieval.

Results can be accessed from the following links:

Google's Universal Sentence Encoder

• Cosine similarity
• Minkowski distance

Baselines

Term Frequency * Inverse Document Frequency (Tf-Idf)

• Corpus as Bag-of-Words (BoW)
• Corpus as BoW with unit vectors
• Corpus as BoW with soft cosine similarity

Latent Semantic Indexing (LSI/LSA)

• Corpus as Bow with 100 topics
• Corpus as Bow with 200 topics
• Corpus as BoW with 200 topics and with unit vectors
• Corpus as Tf-Idf with 100 topics

Random Projections (RP)

• Corpus as Bag-of-Words
• Corpus as Tf-Idf

Latent Dirichlet Allocation (LDA)

• Corpus as Bag-of-Words
• Corpus as Tf-Idf

Hierarchical Dirichlet Process (HDP)

• Corpus as Bag-of-Words
• Corpus as Tf-Idf

Doc2Vec

AppIDs

• With inference
• Without inference

AppIDs and categories

• With inference
• Without inference

AppIDs and genres

• With inference
• Without inference

AppIDs, categories and genres

• With inference
• Without inference

Weighted average of word embeddings

GloVe

Main results

• Cosine similarity
• Minkowski distance

Removing sentence components

• Cosine: remove 1 sentence component
• Cosine: remove 2 sentence components
• Cosine: remove 10 sentence components

Removing word components

• Substract mean from word vectors
• Remove 1 word component
• Remove 2 word components
• Remove 10 word components

Removing both sentence and word components

• Remove 10 sentence components and 10 word components
• Ditto, and replace Tf-Idf weights

Tweaks

• Replace Tf-Idf weights
• Unit vectors
• Non-weighted average
• Without counting words out of vocabulary

Word2Vec

Main results

• Cosine similarity
• Minkowski distance

Cosine: removing sentence components

• Cosine: remove 1 sentence component
• Cosine: remove 2 sentence components
• Cosine: remove 10 sentence components

Minkowski: removing sentence components

• Minkowski: remove 1 sentence component
• Minkowski: remove 2 sentence components
• Minkowski: remove 10 sentence components