Trends, Problems And Solutions of Recommender System

3 minute read

Updated:

Table of Contents

  1. Recommender System at Ground Zero
  2. Breeds of Recommender System
  3. Challenges and Solutions of Recommender System
  4. Link to the Paper

Recommender system

  • Recommender System is an information filtering system that aims at predicting the preference of rating given to an item by any user, thereby helping users make personalised decision.
  • The paper discusses various approaches used in the various recommender systems such as Content based, Collaborative and Hybrid recommender system.
  • It also proposes main challenges and the solutions to these recommendation techniques.

Recommender System at Ground Zero

  • It Emerged in the mid 1990’s, when the researchers started focusing on recommendation problem that depends on the rating method.
  • The aim is to recommend the ‘right’ things to ‘right’ users.
  • Examples: YouTube, LinkedIn, Spotify, Amazon, etc

Breeds of Recommender System

Recommender Description Advantages Disadvantages
Content-based Filtering Systems - Uses information of active users and data about the items.
Steps: 1. Gathers content data about the item( author, cost, i.e. metadata)
2. Process data and extract useful features and elements.
- Doesn’t require data of other users.
- Has capabilities of recommending items to user with unique taste.
Items are limited to their initial descriptions or features.
Collaborative Filtering Systems - Uses information about a set of users and their relations with the item to provide recommendations to the active user.
- Based on a few customers who are most similar to the active users - uses Cosine Similarity.
1. User Based CF: For each user, compute correlation with other users. For each item, aggregate the rating of the users highly correlated with each user.
2. Item-based CF: For each item, compute correlation with other items. For each user, aggregate his rating of the items highly correlated with each item.
Doesn’t need a representation of items. The item can’t be recommended to any user until and unless the item is either rated by another user(s) or correlated with other similar items.
Demographic Filtering Systems - Uses demographic information such as age, gender, education, etc. of people for identifying types of user.
- Uses pre-existing knowledge of demographic information about the users and their opinions
- Doesn’t require history of user ratings.
- Quick, easy and a straightforward method based on few observations
- Recommendations are stereotypical, as it depends on the assumption that users belong to a certain category.
- Security and privacy issues.
Hybrid recommender Systems - Uses a combination of two or more different recommendation techniques.
- Uses both item content and the ratings of all users.
Can overcome various problems caused by a single recommender system Can be complex to implement as the hybridisation method needs to be chosen carefully.

Challenges and Solutions of Recommender System

  • Cold-start
    • Cold-start of new items: When there aren’t enough previous ratings related to the item.
    • Cold-start of new users: When the system don’t have any information related to user’s past purchases or ratings.
    • Solution: Can use demographic information about the user from SNS or the sign-up page. Additionally, an hybrid approach of i.e. using collaborative filtering with demographic recommending approach to suggest items to a new user can be used.
  • Scalability
    • As the number of users and items grows, the system needs more resources and a bigger pool to recommend from.
    • Is an apparent problem in Collaborative filtering approach.
  • Privacy
    • Problem as it collects user’s demographic information - which may breach the privacy of the user.
  • Sparsity
    • Is caused by the insufficient number of user’s interaction with the system and feedback data.
    • Solution: Can be resolved with a hybrid-recommendation. The amount of information people have in common can be increased by using the attributes of an item instead of the item itself.
  • Over-Specialisation (Filter bubble)
    • It gives nothing “surprising” to the user - in other words, users can expect the recommendation they will get.
    • Is an apparent problem in content-based recommendation.
    • Solution: Can be overcome with neighborhood based collaborative filtering technique. K-neighbours can be used to consider the similarity levels between the user/items and candidate neighbours.

Resources

Leave a comment