Other types of

Recommender
Systems
Structural Recommendations in Networks
Example of network based
recommendation
• Networks have become ubiquitous as a modeling tool in many
applications, such as social and information networks. Therefore, it
is particularly useful to discuss various structural elements of a
network that can be recommended in different scenarios.
• Network: a collection of entities that are interconnected with links.
• people that are friends
• computers that are interconnected
• web pages that point to each other
• proteins that interact
• Networks are also called graphs, the entities are nodes, and the
links are edges
Types of Structural
Recommendations
• Recommending nodes by authority and context
• Quality and authority of a node is judged by the incoming links.
• Page-Rank algorithm is adopted by search engines for this purpose. But, it is not personalized.
• Personalized page rank can be used for this purpose
• Recommendation from reputed nodes
• Recommending nodes by example
• Nodes are tagged with similar properties
• Closely related to neighbourhood based approaches
• Target marketing
• Recommending nodes by influence and content
• Nodes with potential to disseminate for information based on their connectivity.
• Social influencer and viral marketing
• Recommending links
• Recommending links to increase the size of a network
• Friend suggestions
Recommending nodes
by authority and context
Recommending nodes by authority
and context
• A health drinks producer wishes to find a brand ambassador for a
newly introduced item. An appropriate person would be whom
everyone respect in this domain. How to find such a person in a social
network?
• Page rank algorithm is the basis
• Initially proposed for ranking Web pages
• achieved using the citation structure of the Web
• A citation can be logically viewed as a vote for the Web page
• To provide a more holistic citation-based vote PageRank is used. The PageRank algorithm
generalizes the notion of citation-based ranking in a recursive way
• Page rank algorithm is not personalized
• Personalized Page Rank
Foundations of page rank algorithm
• In-links of page : The hyperlinks that point to page i from other pages.
• Out-links of page : The hyperlinks that point out to other pages from
page .
• A hyperlink from a page pointing to another page is an implicit
conveyance of authority to the target page. Thus, the more in-links
that a page receives, the more prestige the page has.
• Pages that point to page also have their own prestige scores. A page
with a higher prestige score pointing to is more important than a
page with a lower prestige score pointing to . In other words, a page is
important if it is pointed to by other important pages.
 Foundations of page rank algorithm: Markov Model
• Treat the Web as a directed graph , where V is the set of vertices or
nodes, i.e., the set of all pages, and E is the set of directed edges in the
graph, i.e., hyperlinks.
• Let is the total number of pages and is the number of out-links of page.
• Let be the state transition probability matrix, where

• In case the process satisfies certain property (ergodic, irreducible and

aperiodic), in the long run it converges to a steady state.
 Foundations of page rank algorithm
• The PageRank score of the page (denoted by ) is defined by:

Whereis the number of out-links of page is the total number of

pages
 A is not a stochastic matrix
because the fifth row is all 0
• To convert A to a stochastic transition matrix, add
a complete set of outgoing links from each such
page i to all the pages on the Web.
• Thus the transition probability of going from i to
every page is 1/n assuming uniform probability
distribution.
A is not irreducible. Irreducible means that • That is, we replace each row containing all 0’s
the graph is strongly connected. with e/n, where e is n-dimensional vector of all
1’s.
A directed graph G = (V, E) is strongly connected if and only if, for each pair of nodes u, v
∈ V, there is a path from u to v

A state i is periodic with period k > 1 if k is the smallest number such that all paths leading
from state i back to state i have a length that is a multiple of k. If a state is not periodic (i.e.,
k = 1), it is aperiodic. A Markov chain is aperiodic if all states are aperiodic.
Assumption to deal with the problem:
The random surfer has two options:
1. With probability d, he randomly chooses an out-link to follow.
2. With probability 1d, he jumps to a random page without a link.

With this assumption the improved model is

where E is eeT (e is a column vector of all 1’s) and thus E is a nn square matrix of all 1’s.
1/n is the probability of jumping to a particular page. n is the total number of nodes in the
graph. It is assumed that A has already been made a stochastic matrix.
Using d = 0.9

This matrix can converge to a steady state in the long run. Thus, after simplification
using steady state equations

Or equivalently
The power iteration method for PageRank
Modifications to page rank for
recommending nodes in a
personalized setting
• In personalized page rank algorithm, a personalization vector is
multiplied to the transition probability matrix.
• The personalization vector has one entry per node. If the node is of
interest then the corresponding entry in the vector takes the value 1
otherwise it is zero.
• This enables discovering topic sensitive authoritative nodes.
• The personalized PageRank approach can also be used to discover the
neighborhoods in user-item graphs or user-user graphs in traditional
collaborative filtering applications.
Recommending nodes by example
Finding the nodes with similar
interests
• An manufacturer of golf equipment wishes to target few nodes for
marketing. He must select those node, who are interested in golf. The
interests can be inferred from their own posts, the likes for the others
posts, tagging the related items/news etc.
• Finding such users utilizes the concept of homophily in a social
network, which says, nodes with similar properties are usually
connected.
• Therefore, the profile, properties, and ratings of the neighbourhood
node can be leveraged to make recommendations.
Recommendation by collective
classification
• The actors with specific interest in the network can be
specified with the use of labels. Therefore, a subset of the
nodes are associated with labels.
• It is desired to use these labels as training data to determine
the labels of the other nodes where they are unspecified. It
is assumed that for labeled nodes, the index of the label is
drawn from{1 . . . r}.
• Like the collaborative filtering problem, this is also an
incomplete data estimation problem, except that it is done
in the context of network structures.
Label propagation for classification

Source: velog.io
Challenges
•Convergence is not guaranteed.
•Node feature information was not used.
→ This is because only node labels and network information are used.
 Recommendation by collective
classification
• Because nodes with similar properties are usually connected, it is reasonable to
assume that this is also true of node labels. A solution to this problem is to
examine the k labeled nodes in the proximity of a given node and report the
majority label.
• This approach is, the network analog of a nearest neighbor classifier. However,
such an approach is generally not possible in collective classification because of
the sparsity of node labels.
• In order to handle sparsity, one must not only use the direct connections to
labeled nodes, but also use the indirect connections through unlabeled nodes.
• Two widely discussed algorithms in this regard are:
• Iterative classification algorithm
• Random walk-based method
 Iterative classification algorithm
• Network G = (N,A)
• Class labels: drawn from {1 . . .r}.
• The total number of nodes is denoted by n, from which nt nodes are
unlabeled test nodes.
• Each edge (i, j) ∈ A is associated with the weight wij .
• Node i has two types of features (Content feature, Link feature
• The content Xi is available at the node i in the form of a
multidimensional feature vector.
• ICA algorithm derives a set of link features in addition to the available content features
• A link feature is generated for each class, containing the fraction of its incident nodes
belonging to that class. For each node i, its adjacent node j is weighted by wij for
computing its credit to the relevant class.
Iterative classification algorithm
Recommending nodes by
influence and content
Recommending nodes by influence
and content
• You want to choose few nodes (because you may be budget constrained)
for viral marketing of your product. How to choose few node to ensure
maximum coverage in the network.
• (Influence Maximization) Given a social network G = (N,A), determine a
set of k seed nodes S, influencing which will maximize the overall spread
of influence in the network.
• Each model or heuristic can quantify the influence level of a node with
the use of a function of S that is denoted by f(·). This function maps
subsets of nodes to real numbers representing influence values.
Therefore, after a model has been chosen for quantifying the influence
f(S) of a given set S, the optimization problem is that of determining the
set S that maximizes f(S).
Recommending nodes by influence
and content
• An interesting property of a very large number of influence analysis models is
that the optimized function f(S) is submodular.
• It is a mathematical way of representing the natural law of diminishing
returns, as applied to sets. In other words, if S ⊆ T , then the additional
influence obtained by adding an individual to set T cannot be larger than the
additional influence of adding the same individual to set S.
• Thus, the incremental influence of the same individual diminishes, as larger
supersets of cohorts are available as seeds.
• Two common approaches for defining the influence function f(S) of a set of
nodes S are the Linear Threshold Model and the Independent Cascade Model.
Linear Threshold Model
the algorithm initially starts with an active set of seed nodes S and iteratively
increases the number of active nodes based on the influence of neighboring
active nodes. Active nodes are allowed to influence their neighbors over
multiple iterations throughout the execution of the algorithm until no more
nodes can be activated. The influence of neighboring nodes is quantified with
the use of a linear function of the edge-specific weights bij. For each node i in
the network G = (N, A), the following is assumed to be true:

Each node i is associated with a random threshold θi ∼ U[0,1] that is fixed

up front and stays constant over the course of the algorithm. The total
influence I(i) of the active neighbors of node i on i, at a given time-instant,
is computed as the sum of the weights bij of all active neighbors of i.
 Example of Linear
Threshold Model

(A) node V is activated

and influences W and
U by 0.5 and 0.2,
respectively;
(B) W becomes activated
and influences X and U
by 0.5 and 0.3,
respectively;
(C) U becomes activated
and influences X and Y
by 0.1 and 0.2,
respectively;
(D) X becomes activated
and influences Y by
0.2; no more nodes
can be activated;
https://snap-stanford.github.io/cs224w-notes/network-methods/influence-maximization
process stops.
 Independent Cascade Model
in the independent cascade model, after a node becomes
active, it obtains only a single chance to activate its
neighbors, with propagation probabilities associated with
the edges.
Recommending links
Recommending links
• In many social networks, it is desirable to predict future links between
pairs of nodes in the network.
• For example, commercial social networks, such as Facebook, often
recommend users as potential friends.
• Methods for link prediction
• Neighborhood-Based Measures
• Katz Measure
• Random Walk-Based Measures
• Classification based approach
• Matrix factorization based approach
 Neighborhood-Based Measures
(Common Neighbor Measure) The common-neighbor measure between nodes i and j is
equal to the number of common neighbors between nodes i and j. In other words, if Si is the
neighbor set of node i, and Sj is the neighbor set of node j, the common-neighbor measure is
defined as follows:

The major weakness of the common-neighbor measure is that it

does not account for the relative number of common neighbors
between them as compared to the number of other
connections. It may happen that both the nodes are either
spammers or very popular public figures who were connected to
a large number of other actors. In such a case, the nodes may
have many neighbors in common, just by chance. The Jaccard
measure is designed to normalize for varying degree
distributions.
=4
 Neighborhood-Based Measures
(Jaccard Measure) The Jaccard-based link prediction measure between nodes i and j is
equal to the Jaccard coefficient between their neighbor sets Si and Sj , respectively.

The Jaccard measure adjusts much better to the variations

in the degrees of the nodes between which the link
prediction is measured. However, it does not adjust well to
the degrees of their intermediate neighbors. However, all
of these common neighbors could be very popular public
figures with very high degrees. Therefore, these nodes are
statistically more likely to occur as common neighbors of
many pairs of nodes. This makes them less important in
the link prediction measure.
=4/9
 Neighborhood-Based Measures
(Adamic-Adar Measure) The common-neighbor measure between nodes i and j is
equal to the weighted number of common neighbors between nodes i and j.
=1/log4 + 1/log2 + 1/log2 + 1/log4
Walk-Based Measure
While the neighborhood-based measures provide a robust estimation of the
likelihood of a link forming between a pair of nodes, they are not quite as effective
when the number of shared neighbors between a pair of nodes is small. When
there is significant indirect connectivity through longer paths, walk-based
measures are more appropriate. Katz measure is one metric.

If A is the symmetric adjacency matrix of an undirected network, then the n × n

pairwise Katz coefficient can be computed as
• There are four types of Structural Recommendations: Recommending
nodes by authority and context, Recommending nodes by example,
Recommending nodes by influence and content, and Recommending
links
• Page rank algorithm is the basis for recommending nodes by authority
and context.
• Finding the nodes with similar interests utilizes the concept of homophily
in a social network. Iterative classification algorithm is a classical
approach for this.
• Recommending nodes by influence and content is a problem to select
few seed nodes to maximize the influence. Linear Threshold Model and
the Independent Cascade Model are two classical approaches in this
regard.
• Methods for link prediction include Neighborhood-Based Measures, Katz
Measure, Random Walk-Based Measures, Classification based approach
and Matrix factorization based approach