Abstract:

Yioop is an open-source web portal with search engine, a wiki system and discussion board. Yioop has a recommendation system for recommending groups and discussion threads, but lacks support for recommending media items like books, magazines, videos, etc. This project will extend current recommendation system in Yioop to support recommending media items in PDF, images, videos and e-doc file formats. This will be achieved by completing two main tasks. The first task is to create media job (background process that can either be scheduled or triggered by some event) in Yioop that will update the description of media items by web crawling appropriate web sites depending upon the file format of media items. The second task then aims to use the descriptions of media items for generating Hahs2Vec embeddings word vectors and combine with TF-IDF scores for words to create final vector representation of the media items. Finally find cosine similarity between media items vector and recommend top few media items similar to those which the user has already viewed.

CS297 Results

• Implemented an Emoji Picker tool for the Direct Messages activity in Yioop which allows to use emojis in the text messages
• Developed Yioop UI testing project in Node.js using Selenium web drivers which ensures correct behavior for Yioop UI
• Integrated Stripe APIs for converting unused Yioop credits into real money and transfer them to users bank account
• Understood Hash2Vec word embedding technique and refactored code of former student Anirudh Mallya related to Hash2Vec to generate recommendations for groups and threads in Yioop

Proposed Schedule

Key Deliverables:

• Software
  • Create a Yioop media job which detects media items which are either new or without description and depending upon the file format of media items assign web sources to crawl for finding descriptions
  • Extend above media job, to actually crawl the assigned web sources to find the information about media items based on their titles and update their descriptions in Yioop
  • Update many users experiment in Yioop to create test data about user interaction with media items
  • Extend Yioop recommendation feature to create recommendations for media items using Hash2Vec word embeddings on their descriptions
  • Test and verify the media item recommendations using the test data created by many user experiment
• Report
  • CS 298 Report
  • CS298 Presentation

Innovations and Challenges

• The major challenge for this project is to update the descriptions of media items by crawling information on websites using just their titles
• The innovative aspect of this project is to recommend media types in different file formats using Hash2Vec word embedding and challenging as it is not done in past
• Other challenge for the project is to handle cold start problem where user has never interacted with any media items

References:

[1] Argerich, Luis, Joaquin Torre Zaffaroni, and Matias J. Cano. "Hash2vec, feature hashing for word embeddings." arXiv preprint arXiv:1608.08940 (2016)

[2] Wenga, Carmel, et al. "A Comprehensive Review on Non-Neural Networks Collaborative Filtering Recommendation Systems." arXiv preprint arXiv:2106.10679 (2021)

[3] Y D Pramudita, et al. "Extraction System Web Content Sports New Based On Web Crawler Multi Thread" 2020 J. Phys.: Conf. Ser. 1569 022077

[4] Luong Vuong, Nguyen & Nguyen, Tri-Hai & Jung, Jason. (2020). Content- Based Collaborative Filtering using Word Embedding: A Case Study on Movie Recommendation. 10.1145/3400286.3418253

[5] Zisopoulos, Charilaos & Karagiannidis, Savvas & Demirtsoglou, Georgios & Antaris, Stefanos. (2008). Content-Based Recommendation Systems