CREATING

SUGGESTIONS
AND
RECOMMENDATI
ONS
Unit 3 1
Recommender System
 We ask our friends or relatives for suggestions before making
day-to-day decisions or buying things.
 To making decisions online to buy products, we read reviews
about the products, compare the products' specifications with
other similar products and then we make decisions to buy or
not.
 Because of the growth of information at an exponential rate,
looking for valid information in an online world will be a
challenge. Recommender systems come to our rescue to
provide relevant and required information.
 Some examples of recommender system include suggestions
for products on Amazon, friends' suggestions on social
applications such as Facebook, Twitter, and LinkedIn, video
recommendations on YouTube, news recommendations on
Google News, and so on.
2
Recommender System
 Recommendation engines, a branch of information
retrieval and artificial intelligence, are powerful tools
and techniques to analyze huge volumes of data,
especially product information and user information,
and then provide relevant suggestions based on data
mining approaches.
 A recommendation engine problem is to develop a
mathematical model or objective function which can
predict how much a user will like an item.
 If U = {users}, I = {items} then F = Objective function
and measures the usefulness of item I to user U, given
by:
F:U X I = R, where R = {recommended items}.
 For each user u, we want to choose the item i that 3
Recommender System
 The main goal of recommender systems is to provide
relevant suggestions to online users to make better
decisions from many alternatives available over the
Web.
 A better recommender system is directed more toward
personalized recommendations by taking into
consideration the available digital footprint of the user,
such as user-demographic information, transaction
details, interaction logs, and information about a
product, such as specifications, feedback from users,
comparison with other products, and so on, before
making recommendations.
 The most common recommender systems are:
collaborative filtering recommender system and 4
content-based recommender system.
Collaborative Filtering Recommender Systems
 Collaborative filtering recommender systems are basic forms of
recommendation engines.
 In this type of recommendation engine, filtering items from a
large set of alternatives is done collaboratively by users'
preferences.
 The basic assumption in a collaborative filtering recommender
system is that if two users shared the same interests as each
other in the past, they will also have similar tastes in the future.
 If, for example, user A and user B have similar movie
preferences, and user A recently watched Titanic, which user B
has not yet seen, then the idea is to recommend this unseen
new movie to user B.
 The movie recommendations on Netflix are one good example
of this type of recommender system.

5
Collaborative Filtering Recommender Systems
 The idea is very simple: given the ratings of a user towards items,
find all the users similar to the active user who had similar ratings in
the past and then make predictions regarding all unknown items
that the active user has not rated but are being rated in their
neighborhood.
 In these types of systems, the main actors are the users, products,
and user’s rating information such as continuous, interval-based,
ordinal, binary, and unary towards the products.
 The two common collaborative filtering recommender
systems are: User-based collaborative filtering
and Item-based collaborative filtering. These two
are also called memory-based or neighborhood-
based collaborative filtering.
6
Collaborative Filtering Recommender Systems
 User-based collaborative filtering:
In this case, the ratings provided by like-minded users of a target user A
are used in order to make the recommendations for A.
The basic idea is to determine users, who are similar to the target user A,
and recommend ratings for the unobserved ratings of A by computing
weighted averages of the ratings of this peer group.
Therefore, if user A and user B have rated movies in a similar way in the
past, then one can use user A’s observed ratings on the movie M to predict
B’s unobserved ratings on this movie.
In general, the k most similar users to user B can be used to make rating
predictions for user B.
Similarity functions (cosine similarity, Pearson correlation coefficient etc.)
are computed between the rows of the ratings matrix to discover similar
users.
7
Collaborative Filtering Recommender Systems
Example 1: Suppose, the ratings of five users U1, U2, U3, U4, and U5 are
indicated for six items denoted by 1, 2, 3, 4, 5, and 6. Each rating is
drawn from the range {1 ... 7}. Find the predictions of user U3 for items 1
and 6 on the basis of the ratings in the table below. Use Pearson
correlation coefficient to find similarity between users and consider top-2
closest users.
Sim(x, y) = Pearson(x, y) =

User/ I1 I2 I3 I4 I5 I6
Item
U1 7 6 7 4 5 4
U2 6 7 ? 4 3 4
U3 ? 3 3 1 1 ?
U4 1 2 2 3 3 4
U5 1 ? 1 2 3 3
8
Collaborative Filtering Recommender Systems
Solution: The first step is to compute the similarity between user U3 and all the other
users.
Pearson(U1, U3) = = 0 .89442719
Similarly, we can calculate,
Pearson(U2, U3) = 0.97072534
Pearson(U4, U3) = -1
Pearson(U5, U3) = -0.8660254
Hence, the top-2 closest users to user U3 are users U1 and U2 according to Pearson
correlation coefficient. By using the Pearson-weighted average of the raw ratings of users
U1 and U2, the following predictions are obtained for user U3 with respect to her unrated
items 1 and 6.
Rank(U3, 1) = (7 * 0.89442719 + 6 * 0.97072534) / (0.89442719 + 0.97072534) = 6.48
Rank(U3, 6) = (4 * 0.89442719 + 4 * 0.97072534) / (0.89442719 + 0.97072534) = 4

9
Collaborative Filtering Recommender Systems
Exercise 1: Suppose, the ratings of five users U1, U2, U3, U4, and U5 are
indicated for six items denoted by 1, 2, 3, 4, 5, and 6. Each rating is
drawn from the range {1 ... 7}. Find the predictions of user U3 for items 1
and 6 on the basis of the ratings in the table below. Use cosine similarity
to find similarity between users and consider top-2 closest users.
Sim(x, y) = Cosine(x, y) =

User/ I1 I2 I3 I4 I5 I6
Item
U1 7 6 7 4 5 4
U2 6 7 ? 4 3 4
U3 ? 3 3 1 1 ?
U4 1 2 2 3 3 4
U5 1 ? 1 2 3 3

10
Collaborative Filtering Recommender Systems
 Item-based collaborative filtering:
In order to make the rating predictions for target item B by user A, the first
step is to determine a set S of items that are most similar to target item B.
The ratings in item set S, which are specified by A, are used to predict
whether the user A will like item B.
In general, the k most similar items to item B can be used to make rating
predictions for item B.
Similarity functions are computed between the columns of the ratings
matrix to discover similar items.

11
Collaborative Filtering Recommender Systems
Example 2: Suppose, the ratings of five users U1, U2, U3, U4, and U5 are
indicated for six items denoted by 1, 2, 3, 4, 5, and 6. Each rating is
drawn from the range {1 ... 7}. Find the predictions of item 1 for user U3
on the basis of the ratings in the table below. Use Pearson correlation
coefficient to find similarity between users and consider top-2 closest
items.
User/ I1 I2 I3 I4 I5 I6
Item
U1 7 6 7 4 5 4
U2 6 7 ? 4 3 4
U3 ? 3 3 1 1 ?
U4 1 2 2 3 3 4
U5 1 ? 1 2 3 3

12
Collaborative Filtering Recommender Systems
Solution: The first step is to compute the similarity between item 1 and all the other
items.
Pearson(1, 2) = 0.94063416
Pearson(1, 3) = 0.98782916
Pearson(1, 4) = 0.89714996
Pearson(1, 5) = 0.67675297
Pearson(1, 6) = 0.57263713
Hence, the top-2 closest items to item 1 are 2 and 3. By using the Pearson-weighted
average of the raw ratings of user U3 for items 2 and 3 is used to predict the rating of
item 1 for user U3.
Rank(U3, 1) = (3 * 0.94063416 + 3 * 0.98782916) / (0.94063416 + 0.98782916) = 3

13
Collaborative Filtering Recommender Systems
 Advantages:
Easy to implement.
Neither the content information of the products nor the users' profile
information is required for building recommendations.
New items are recommended to users giving a surprise factor to the users.
 Disadvantages:
This approach is computationally expensive as all the user, product, and
rating information is loaded into the memory for similarity calculations.
This approach fails for new users where we do not have any information
about the users. This problem is called the cold-start problem.
This approach performs very poorly if we have little data.
Since we do not have content information about users or products, we
cannot generate recommendations accurately based on rating information
only.
14
Content-based Recommender Systems
 In real life, recommendation should be based on contents of items,
such as, genre, actor, director, story, and screenplay in case of
movies.
 A recommendation that is targeted at a personalized level and that
considers individual preferences and contents of the items for
generating recommendations is called a content-based
recommender system.
 Content-based recommendation engines solve the cold-start
problem that new users face in the collaborative filtering approach.
When a new user comes, based on the preferences of the person
we can suggest new items that are similar to their tastes.

15
Content-based Recommender Systems
 Building content-based recommender systems involves three main
steps, as follows:
1. Generating content information for items (item profile generation)
2. Generating a user profile and preferences with respect to the features of
the items (user profile generation)
3. Generating recommendations and predicting a list of items to the user
 Item Profile Generation:
In this step, we extract the features that represent the item. Most commonly
the content of the items is represented in the vector space model with item
names as rows and features as columns.
We extract relevant features and their relative importance score associated
with the item. These scores may be Boolean or real valued.
In the case of movies, item profile might be a Boolean vector where 1
represents presence of feature and 0 represents absence of feature.
16
Content-based Recommender Systems
In case of text documents, item profile might be a real valued vector. Item
profile is generated using term frequency inverse document frequency (tf-
idf). Tf-idf shows the feature relative importance associated with the item.
 User Profile Generation:
In this step, we build the user profile matching the item feature using the
utility matrix representing the connection between users and items.
The utility matrix entries could be just 1’s representing user purchases or a
similar connection, or they could be arbitrary numbers representing a rating
or degree of affection that the user has for the item.
 Generating Recommendation:
With profile vectors for both users and items, we can estimate the degree to
which a user would prefer an item by computing the cosine similarity
between user’s and item’s vectors.
The greater the cosine angle the more likely the user is to like a movie and
hence it can be recommended to the user. 17
Content-based Recommender Systems
 Example: Find the predictions of Movie M4 for user U1 on the basis of the
information given in the table below.
Movie/ U U U U U U Genre
User 1 2 3 4 5 6
M1 1 1 ? 1 1 ? Romanc
e
M2 1 1 1 1 1 ? Thriller
M3 1 1 1 1 1 1 Action
M4 ? 1 1 1 1 1 Romanc
e
M5 1 1 1 1 1 ? Crime
M6 1 1 1 1 1 1 Crime

18
Content-based Recommender Systems
Item Profile Generation:
Movie/ Romanc Thriller Actio Crime
Genre e n
M1 1 0 0 0
M2 0 1 0 0
M3 0 0 1 0
M4 1 0 0 0
M5 0 0 0 1
M6 0 0 0 1

19
Content-based Recommender Systems
User Profile Generation:
Movie/ Romanc Thriller Actio Crime
Genre e n
U1 0.2 0.2 0.2 0.4
U2 0.33 0.17 0.17 0.33
U3 0.2 0.2 0.2 0.4
U4 0.33 0.17 0.17 0.33
U5 0.33 0.17 0.17 0.33
U6 0.33 0 0.33 0.33

Sim(M4, U1) = Cosine(M4, U1) = 0.38

20
Mining Frequent Patterns
 Apriori
 FP growth

21