Research Paper Recommendation System

[Mikeyzgoat]

Deep Learning Simplified Repository (Proposing new issue)

🔴 Project Title : Research Paper Recommendation System

🔴 Aim : To create a recommendation system to suggest research papers based on user input (input format can be decided by the candidate)

🔴 Dataset : https://www.kaggle.com/datasets/sumitm004/arxiv-scientific-research-papers-dataset

🔴 Approach : Use NLP, ML algorithms such as BERT, FAISS (similarity search) and other classification algorithms to narrow-down on the best research papers based on user preferences.

---

🔴🟡 Points to Note :

• The issues will be assigned on a first come first serve basis, 1 Issue == 1 PR.
• "Issue Title" and "PR Title should be the same. Include issue number along with it.
• Follow Contributing Guidelines & Code of Conduct before start Contributing.

---

✅ To be Mentioned while taking the issue :

• Full name :
• GitHub Profile Link :
• Email ID :
• Participant ID (if applicable):
• Approach for this Project :
• What is your participant role? (Mention the Open Source program)

---

Happy Contributing 🚀

All the best. Enjoy your open source journey ahead. 😎

[github-actions]

Thank you for creating this issue! We'll look into it as soon as possible. Your contributions are highly appreciated! 😊

[Mikeyzgoat]

@abhisheks008 can you add the tags to this please. Place it under CODEPEAK 25. Thank you

[Balaji1810psna]

There are many research paper website soruces
Example :IEEE Explore, Google scholar etc.
If you want search that website

Google scholar link:
https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://scholar.google.com/&ved=2ahUKEwiWjrfKzNKPAxXezjgGHQdlEKgQFnoECBAQAQ&usg=AOvVaw0EsBbxLkLeagBX_vue5ltw

[Anshika0998]

Full name : Anshika shukla

GitHub Profile Link : https://github.com/Anshika0998

Email ID : @shuklaanshika0998@gmail.com

Participant ID (if applicable): @Anshika0998

Approach for this Project :
I plan to build a Research Paper Recommendation System using NLP and semantic similarity techniques. The workflow would include:

• Data preprocessing and cleaning of research paper metadata such as title, abstract, authors, and categories

• Generating contextual embeddings using transformer-based models like BERT/Sentence-BERT

• Using FAISS for efficient similarity search over high-dimensional embeddings

• Designing a recommendation pipeline that suggests relevant papers based on:

  • paper title
  • abstract
  • keywords
  • user interests/preferences

• Ranking recommendations using cosine similarity and relevance scoring

• Building a simple and user-friendly interface/API for querying recommendations

• Evaluating recommendation quality using similarity metrics and qualitative testing

I also plan to keep the architecture modular so that newer embedding models or ranking methods can be integrated later.

Participant - GSSoC 2026

[parm-08]

Full Name: Parmpreet Kaur

GitHub Profile Link: https://github.com/parm-08

Email ID: parmpreetkaur2005@gmail.com

Participant ID (if applicable): parm-08

Approach for this Project:

I would like to work on developing a Research Paper Recommendation System using NLP and semantic similarity techniques. My approach includes preprocessing and cleaning the dataset, generating contextual embeddings using transformer-based models such as BERT/Sentence-BERT, and implementing FAISS for efficient similarity search.

The recommendation system will suggest relevant research papers based on paper title, abstract, keywords, categories, and user preferences. I also plan to implement cosine similarity-based ranking to improve recommendation quality and create a modular architecture for future scalability and improvements.

Additionally, I will focus on:
• Exploratory Data Analysis (EDA)
• Efficient recommendation pipeline design
• Proper documentation and clean code structure
• User-friendly interface/API integration
• Performance evaluation and testing

Skills & Technologies:
Python, NLP, Machine Learning, Scikit-learn, Transformers, FAISS, Pandas, NumPy, basic backend integration, and data preprocessing.

I am excited to contribute to this project, collaborate with mentors and contributors, and enhance the recommendation system with scalable and efficient solutions.

Participant Role: GSSoC 2026 Contributor

[Debarati-06-ai]

I would like to work on this as
a GSSoC 2026 contributor.
@abhisheks008 please assign this to me!

[abhisheks008]

Hi @Anshika0998, @parm-08 & @Debarati-06-ai more or less you all are suggesting same suggestion for the models. As per the project guidelines for every project in this repository we have implemented 3-4 models. Can you please brainstorm a bit and share your thoughts on the models those you are going to implement with this dataset and why? What will be the accuracy for each of the models?

[Debarati-06-ai]

I brainstormed some different models that can be compared in this project:

1. TF-IDF + Cosine Similarity
   Basic NLP model using keyword matching.
   Fast and simple baseline model.
2. Word2Vec/Doc2Vec
   Understands semantic meaning better than TF-IDF.
   Gives better recommendations for related topics.
3. BERT/Sentence-BERT + FAISS
   Main deep learning model.
   Understands context of research paper abstracts.
   FAISS helps in fast similarity search for large datasets.
   Expected to give the best recommendation accuracy.
4. Hybrid Model
   Combines semantic search with filters like category/author/year.
   Can improve recommendation quality further.

[abhisheks008]

I brainstormed some different models that can be compared in this project:

1. TF-IDF + Cosine Similarity
   Basic NLP model using keyword matching.
   Fast and simple baseline model.
2. Word2Vec/Doc2Vec
   Understands semantic meaning better than TF-IDF.
   Gives better recommendations for related topics.
3. BERT/Sentence-BERT + FAISS
   Main deep learning model.
   Understands context of research paper abstracts.
   FAISS helps in fast similarity search for large datasets.
   Expected to give the best recommendation accuracy.
4. Hybrid Model
   Combines semantic search with filters like category/author/year.
   Can improve recommendation quality further.

• How will you test the models?
• What will be the metrices for the checking the accuracy of the models?
• Why you have chosen these models? Did you find something in the dataset?

[Debarati-06-ai]

The models can be tested by splitting the dataset into train and test data and checking how relevant the recommended papers are for a given input paper/query.
Metrics for evaluation:
Precision@K : checks how many recommended papers are relevant.
Recall@K : checks how many relevant papers were successfully recommended.
Cosine similarity score : measures similarity between papers.
Response time : compares speed of different models.
Reason specifically why I chose these:
TF-IDF is a simple baseline model for keyword matching.
Word2Vec/Doc2Vec captures semantic similarity better than traditional NLP methods.
BERT/Sentence-BERT understands context and meaning of abstracts more effectively.
FAISS is useful because the dataset is large and similarity search needs to be fast.

[parm-08]

Hi @Anshika0998, @parm-08 & @Debarati-06-ai more or less you all are suggesting same suggestion for the models. As per the project guidelines for every project in this repository we have implemented 3-4 models. Can you please brainstorm a bit and share your thoughts on the models those you are going to implement with this dataset and why? What will be the accuracy for each of the models?

Hi @abhisheks008, thank you for the feedback and suggestion to brainstorm further.
After analyzing the dataset and project requirements more deeply, I refined my approach to include multiple recommendation strategies instead of using only similar baseline models.

I am planning to implement:

TF-IDF + Cosine Similarity for keyword-based recommendations
BERT embeddings for semantic understanding of research papers
FAISS for efficient large-scale similarity search
Random Forest / XGBoost for topic classification and filtered recommendations

I chose these models because each solves a different part of the recommendation pipeline:

TF-IDF gives a strong lightweight baseline
BERT improves semantic recommendation quality
FAISS makes retrieval scalable and faster
Classification models help narrow recommendations based on domains/topics

Expected performance (approximate after preprocessing and tuning):

TF-IDF: ~70–80% relevance
BERT: ~85–92% semantic relevance
FAISS: ~85–90% retrieval relevance
Random Forest: ~80–85% accuracy
XGBoost: ~85–90% accuracy

I will also compare all models using:

Precision@K
Recall@K
F1-score
Cosine similarity metrics
Classification accuracy

along with detailed evaluation and comparison tables in the PR.

I’d really love to contribute to this project and implement a complete hybrid recommendation pipeline with proper NLP preprocessing and deployment. Thank you!

[abhisheks008]

Based on the comments, this issue is assigned to @parm-08 for the clarity in the comments.