What Is Recommendation System?

• A recommendation system is a subclass of Information filtering

Recommender Systems Systems that seeks to predict the rating or the preference a user
might give to an item.
• In simple words, it is an algorithm that suggests relevant items to
Dr. Rabins Porwal users.
• Ex.: In the case of Netflix which movie to watch, In the case of e-commerce
which product to buy, or In the case of kindle which book to read, etc.

Dr. Rabins Porwal 2

1 2

Use-
Use-Cases of Recommender System Types/ Approaches of Recommender Systems
• There are many use-cases of it. Some are • Content-based filtering
A. Personalized Content: Helps to improve the on-site experience by creating • Collaborative filtering
dynamic recommendations for different kinds of audiences like Netflix
does. • Session-based recommender systems
B. Better Product search experience: Helps to categorize the product based • Reinforcement learning for recommender systems
on their features. Ex.: Material, Season, etc.
• Multi-criteria recommender systems
• Risk-aware recommender systems
• Mobile recommender systems
• Hybrid recommender systems

Dr. Rabins Porwal 3 Dr. Rabins Porwal 4

3 4

1. Content-
Content-Based Filtering
Example (Movie Recommendation System):
• In this type of recommendation
system, relevant items are • In movie recommendation system, every movie is associated with its
shown using the content of the genres which is referred to as tag/attributes.
previously searched items by
the users. Here content refers • Assume user A comes and initially system doesn’t have any data
to the attribute/tag of the about user A. so initially, the system tries to recommend the popular
product that the user like. movies to the users, or the system tries to get some information of
• In this type of system, products the user by getting a form filled by the user. After some time, users
are tagged using certain might have given a rating to some of the movies like it gives a good
keywords, then the system tries
to understand what the user rating to movies based on the action genre and a bad rating to the
wants, and it looks in its movies based on the anime genre. So here system recommends
database and finally tries to action movies to the users.
recommend different products • But here it can’t be said that the user dislikes animation movies because may
that the user wants. be the user dislikes that movie due to some other reason like acting or story
but actually likes animation movies and needs more data in this case.
Dr. Rabins Porwal 5 Dr. Rabins Porwal 6

5 6

Compiled by: Dr. Rabins Porwal 1

 Advantages Disadvantages
• Model doesn’t need data of other users since recommendations are • Feature representation of items is hand-engineered to some extent,
specific to a single user. this tech requires a lot of domain knowledge.
• It makes it easier to scale to a large number of users. • The model can only make recommendations based on the existing
• The model can capture specific interests of the user and can interest of a user. In other words, the model has limited ability to
recommend items that very few other users are interested in. expand on the user’s existing interests.

Dr. Rabins Porwal 7 Dr. Rabins Porwal 8

7 8

2. Collaborative Based Filtering

Contd…
Contd…
• Recommending the new items to users based on the interest and
preference of other similar users is basically collaborative-based • This overcomes the disadvantage of content-based filtering as it will
filtering. Ex.: When we shop on Amazon, it recommends new products use the user-interaction instead of content from the items used by
saying “Customer who bought this also bought” as shown below: the users. For this, it only needs the historical performance of the
users. Based on the historical data, with the assumption that user,
who has agreed in past, tends to agree in future too.
• There are 2 types of collaborative filtering:
• User-based collaborative filtering
• Item-based collaborative filtering

Dr. Rabins Porwal 9 Dr. Rabins Porwal 10

9 10

A. User-
User-Based Collaborative Filtering B. Item-
Item-Based Collaborative Filtering
• Rating of the item is done using the rating of • The rating of the item is predicted using the
neighbouring users. In simple words, It is based user’s own rating on neighbouring items. In
on the notion of users’ similarity. simple words, it is based on the notion of
• Example: In the adjacent diagram, there are 3 item similarity.
children named A, B, C, and 4 fruits, i.e., grapes,
strawberry, watermelon, and orange • Example: Here the only difference is that we
respectively. Based on the image let A see similar items, not similar users like if you
purchased all 4 fruits, B purchased only see grapes and watermelon, you will realize
strawberry and C purchased strawberry as well that watermelon is purchased by all of them,
as watermelon. Here A & C are similar kinds of but grapes are purchased by Children A & B.
users, because of this C will be recommended Hence Child C is recommended grapes.
grapes and orange as shown by dotted line.

Dr. Rabins Porwal 11 Dr. Rabins Porwal 12

11 12

Compiled by: Dr. Rabins Porwal 2

 Contd…
Contd… Advantages
• Now after understanding both collaborative filtering, you may be • It works well even if the data is small.
wondering which to use when. • This model helps the users to discover a new interest in a given item
• If No. of items is greater than No. of users, go with user-based collaborative but the model might still recommend it because similar users are
filtering as it will reduce the computation power and if No. of users is greater
than No. of items, go with item-based collaborative filtering.
interested in that item.
• Example: Amazon has lakhs of items to sell but has billions of customers. • No need of domain Knowledge.
Hence Amazon uses item-based collaborative filtering because of less no. of
products as compared to its customers.

Dr. Rabins Porwal 13 Dr. Rabins Porwal 14

13 14

Evaluation Metrics
Disadvantages
• How to evaluate whether the given model is recommending the right things
or not and how many relevant things this system predicts: here comes
• It cannot handle new items because the model doesn’t get trained on evaluation metrics. There are several metrics for evaluating the model but 4
the newly added items in the database. This problem is known as major metrics are given below:
Cold Start Problem. • 1. Mean Average precision at K
• It gives how much relevant is the list of recommended items. Here precision at K
• Side Features don’t have much importance. Here Side features can be means recommended items are in top K sets that are relevant.
actor name or releasing year in the context of movie • 2. Coverage
recommendation. • It is the percentage of items in the training data model to recommend in test sets. Or
Simply, the percentage of a possible recommendation system can predict.
• 3. Personalization
• It is basically how many same items the model recommends to different users. Or, the
dissimilarity between users’ lists and recommendations.
• 4. Intralist Similarity
• It is an average cosine similarity of all items in a list of recommendations.
Dr. Rabins Porwal 15 Dr. Rabins Porwal 16

15 16

Compiled by: Dr. Rabins Porwal 3