Ideas

Ideas

Created: 2022-03-28 10:42
#note

My RecSys could work in the following way:

• documents and pictures are embedded using Data2Vec - A General Framework for Self-supervised Learning in Speech, Vision and Language;
• User embedding is obtained using data2vec on the history of liked and/or created posts;
• topic modeling (Top2Vec or BERTopic) is used to extract the most important features of the documents, i.e. the topics. Do the same for the user embedding;
• Extract the recommended posts using FAISS or NMSLib;
• Rank the pictures of the posts.

I need to compare results of two scenarios:

• using topic modeling to reduce number of features;
• without using topic modeling.

References

1. LDA and LSA for recommender systems
2. Paper recommendation
3. ANN for RecSys

Code

Tags

#to-do #ideas