SIGNAL PROCESSING TECHNIQUES FOR INTERPOLATION IN GRAPH STRUCTURED

DATA

Sunil K Narang, Akshay Gadde and Antonio Ortega

Ming Hsieh Department of Electrical Engineering

University of Southern California
[email protected], [email protected],[email protected]

ABSTRACT protein network, one relies on some notions of ‘closeness’ or ‘dis-

tance’ among the nodes. In other words, the graph functions of inter-
In this paper, we propose a novel algorithm to interpolate data de- est are slowly-varying or “smooth” on the graph. Smoothness is well
fined on graphs, using signal processing concepts. The interpo- studied in the classical signal processing domain, where it is mea-
lation of missing values from known samples appears in various sured in terms of the frequency contents of a signal. For example,
applications, such as matrix/vector completion, sampling of high- there are well known sampling results for recovering smooth ban-
dimensional data, semi-supervised learning etc. In this paper, we dlimited signals from only a few of their samples via interpolation.
formulate the data interpolation problem as a signal reconstruction In this paper, we focus on the development of similar interpolation
problem on a graph, where a graph signal is defined as the infor- techniques for graph signals.
mation attached to each node (scalar or vector values mapped to the
set of vertices/edges of the graph). We use recent results for sam- The main objective in designing an interpolation method is of
pling in graphs to find classes of bandlimited (BL) graph signals that course achieving high accuracy. A major challenge in achieving this
can be reconstructed from their partially observed samples. The in- goal is managing to do so with reasonable computational complexity.
terpolated signal is obtained by projecting the input signal into the Graph based interpolation approaches can be broadly divided into
appropriate BL graph signal space. Additionally, we impose a ‘bilat- two categories: a) local methods and b) global methods. In local in-
eral’ weighting scheme on the links between known samples, which terpolation methods, such as kNN methods, the predicted value at an
further improves accuracy. We use our proposed method for col- unknown node is computed as a weighted combination of k-nearest
laborative filtering in recommendation systems. Preliminary results known samples [8]. Although the method is computationally simple
show a very favorable trade-off between accuracy and complexity, and efficient, it does not capture global information, as well as the
compared to state of the art algorithms. dependencies that exist between known samples. On the other hand,
global methods, such as [9, 10], predict the value of all unknown
Index Terms— Graph signal processing, sampling in graphs, nodes at once, by selecting as solution a function that matches the
spectral graph theory, recommendation systems values at known nodes while satisfying certain global ”smoothness”
conditions. Global methods are computationally more expensive but
provide more accurate results. However, it is not clear how to opti-
1. INTRODUCTION
mize the choice of objective function. Although our approach leads
to a similar least-squares based interpolation as in [9], the signal
Graphs are a natural tool to represent data in many domains such as
processing perspective allows us to choose an optimal objective cri-
protein interaction networks [1], recommendation systems [2] and
terion, which will be shown to lead to better interpolation. Another
social networks [3]. A graph, represented as G = (V, E), consists
problem is the choice of graph to interpolate data upon. The ex-
of a set of nodes V and a set of links E connecting these nodes. The
ample in Figure 1 demonstrates the choice of either operating on a
data on the graphs is often represented as a scalar or vector valued
bare-bones star graph or a complete graph with all connections be-
function attached to the vertices of the graph (see [4] for a com-
tween movies. Interpolation can be defined on both graphs, so it is
prehensive review). An important area of research is the interpola-
unclear which of the two (or any other graph) should be chosen.
tion problem in graph structured data. This arises in many guises,
such as in semi-supervised learning of categorical data (see [5]), In this paper, we take inspiration from signal processing
ranking problems [6], and missing value prediction such as matrix- techniques to formulate the partially known graph function as a
completion problems [7]. A common theme in all these applications downsampled-upsampled (DU) signal. In the regular signal domain,
is that the goal is to predict the property of some nodes (class, rank- the original signal is recovered from its DU signal by applying a
ing or function), by interpolating the property values from a known low-pass filter. Similarly, in the graph domain, we design low-pass
set of nodes. The accuracy of all linear and non-linear interpolation filters to recover the original graph signal from its DU signal. The
methods on graphs rely on the implicit assumption that nodes close proposed approach is based on recent results for sampling in ban-
to each other (in terms of the similarity captured by link-weights in dlimited graph signals by Pesenson [11] and our own prior work
the graph) would usually have similar signal values. For example, in [12]. In this approach, given a graph and the set of nodes for
in an item-item graph in a recommendation system, a typical user which the data is known we compute an optimal cut-off frequency,
would rate two similar items with similar ratings. In the same way, such that the reconstruction is exact if the original graph signal is
when predicting the functions of unannotated proteins based on a bandlimited to this frequency, and a stable reconstruction is obtained
for non-bandlimited signals. The errors can be further minimized by
This work was supported in part by NSF under grant CCF-1018977. increasing smoothness of the function using a “bilateral- weighting”

978-1-4799-0356-6/13/$31.00 ©2013 IEEE 5445 ICASSP 2013

 complete graph

Unrated
kNN graph
Unrated
demonstrate increasing oscillatory behavior as the magnitude of the
vertex vertex
graph frequency increases (see [13] for details). The graph Fourier
transform (GFT) of a signal f is defined as its projection onto the
eigenvectors of the graph, i.e., f˜(λi ) =< f , ui >, or in matrix
form f̃ = Ut f .

0.11789 0.75042 0.11789 1.1475 2.2. Sampling Theory for Graph Signals
(a) (b)
We start by revisiting the theory of downsampling graph signals in
Fig. 1: An instance of predicting ratings of an unknown movie node (in red) the recent work by Pesenson [11], as well as its links to our prior
using ratings of a known set of movie nodes (in blue), in MovieLens 100k work in [12]. A signal is said to be bandlimited to the graph fre-
dataset: (a) star graph commonly used in kNN [8] prediction methods, quency band [0, ω) on a graph G, if its GFT has support only at
which ignores all the links between known movie nodes. (b) alternative frequencies [0, ω). The space of ω-bandlimited signals is called the
graph that contains the star graphs and all the links between movies in the Paley-Wiener space and is denoted by P Wω (G). It is easy to prove
known set of movies. that if f ∈ P Wω (G), then:

kLf k ≤ ωkf k (1)

step on the underlying graphs. To the best of our knowledge, this is
a novel perspective for addressing graph based interpolation prob- where k.k denotes the l2 norm. The theory presented by Pesen-
lems, and our preliminary results indicate a promising advantage son [11], describes a method to reconstruct any signal in P Wω (G)
over existing methods. The rest of the paper is organized as follows: space from a special subset of nodes called the uniqueness set. This
in Section 2, we introduce the basic theory needed to understand the is defined as follows:
proposed work, in Section 3, we describe our proposed interpolation
method. In Section 4, we demonstrate our proposed method on a real Definition 1 (Uniqueness Set) A subset of nodes S ⊂ V is a
movie dataset, and finally in Section 5, we conclude and describe uniqueness set for a space P Wω (G), ω > 0, if for any two signals
future work. from P Wω (G), the fact that they coincide on S implies that they
coincide on V.
2. BASIC THEORY
The above definition implies that it is sufficient to know the value
2.1. Notations of a ω-bandlimited graph signal only on the uniqueness set S. The
uniqueness set is in turn found by finding its complement set S c ,
An undirected simple graph G = (V, E) is a collection of nodes which is a Λ-set with Λ = 1/ω. A Λ-set is defined as follows:
V = {1, 2, ...N } connected together by set of links E = {(i, j, wij )},
i, j ∈ V where (i, j, wij ) denotes the link of weight wij be- Definition 2 (Λ-set) A subset of nodes Q ⊂ V is a Λ-set if any
tween node i and j. The adjacency matrix W of the graph is φ ∈ L2 (Q) admits a Poincaré inequality with a constant Λ > 0,
an N × N matrix such that W (i, j) = wij , the degree di a i.e.,
node i is the sum of link-weights connected to node i, the de- kφk ≤ ΛkLφk, φ ∈ L2 (Q) (2)
gree matrix D = diag{d1 , d2 , ..., dN } is a diagonal matrix, and
the combinatorial Laplacian matrix is L = D − W. We use where L2 (Q) is the space of all graph signals that are zero every-
the normalized form of adjacency matrix W = D−1/2 WD−1/2 where except at the subset of nodes Q ⊂ V.
and Laplacian matrix L = D−1/2 LD−1/2 = I − W. L is a
The downsampling results given by Pesenson [11, Theorem 1.2 and
positive semi-definite matrix, and has an orthogonal set of eigen-
4.1] are formally:
vectors denoted U = {u1 , u2 , ...un } corresponding to eigenvalues
σ(G) = {λ1 , λ2 , ..., λN }, respectively. We denote < a, b >= at b Theorem 1 A set S ⊂ V is a uniqueness set for any signal in space
as the inner-product of vectors a and b, where at is the transpose of P Wω (G), if its complement S c is a Λ-set with 0 < ω < 1/Λ,
a, and S c = V − S, the complement set of S on graph G. Define which implies that there exists a frame {Θs }s∈S in the subspace
(A)S1 ,S2 to be the submatrix of any matrix A corresponding to P Wω (G) such that the following reconstruction formula holds true
rows corresponding to S1 and columns corresponding to S2 . Simply for all signals f ∈ P Wω (G):
denote (A)S,S = (A)S for brevity. We denote 0 and 1 as the all
zero and all one functions of size N . In order to use signal pro- X
f (v) = f (s)Θs (v), v ∈ V (3)
cessing tools, we define a graph function as a scalar valued discrete
s∈S
signal f : V → R, such that f (i) is the value of the sample at node i.
The downsampling-upsampling (DU) operation on the graph signal
is defined as discarding the values of signal not on a selected subset Theorem 1 provides a sufficient condition for reconstructing any sig-
of nodes (say S) and replacing them with 0. By an appropriate nal in P Wω (G) from its DU signal. The reconstruction method
permutation, let us choose an indexing such that S = {1, 2, ...|S|}, is described in Section 3. A corollary of Theorem 1 provides the
and S c = {|S| + 1, |S| + 2, ..., N }. Then the DU signal is denoted uniqueness set for the special case when ω = 1 [11, Lemma 3.5].
as fS = [f (S)t 0(S c )t ]t .
Similar to classical Fourier transform, the eigenvectors and Corollary 1 Any vertex cover 1 S ⊂ V of the graph G is a unique-
eigenvalues of the Laplacian matrix L provide a spectral interpreta- ness set for all signals f ∈ P Wω (G) with ω = 1.
tion of the graph signals. The eigenvalues {λ1 , λ2 , ..., λN } can be
treated as graph frequencies, and are always situated in the interval 1 A vertex cover for a graph G is a subset S ⊂ V such that each edge of
[0, 2] on the real line. The eigenvectors of the Laplacian matrix the graph is incident on at least one vertex in S.

5446
This result is also consistent with the result in our prior work on bi- for piecewise constant signals, a common practice is to find LS ap-
partite graphs in [12, 14]. Thus, an interpolation on a graph can be proximation of function g = D1/2 f . Thus, we want to compute
posed as the problem of first defining the set of nodes with known interpolated signal ĝ such that (ĝ)S = (g)S and:
sample values as a uniqueness set S, then identifying the maximum
KΛ
ω such that S is a uniqueness set for signals in P Wω (G), and then X
reconstructing the signal values on the complement set S c by us- ĝ = x(k)uλk = UK ∗ x, (6)
k=1
ing (3).
While Pesenson’s work proves the existence of such ω, it does where x(k) = g̃(λk ) is the kth GFT coefficient of g, and x =
not provide a method to compute it, and only considers how to in- [x1 , x2 , ...xKΛ ]. Comparing only the known set S on both sides
terpolate unknown values (in S c ) in the case when the signal is ω- in (6), we obtain a linear system of equations: g(S) = (UK ∗ )S x.
bandlimited. In the next section, we provide a result to compute Theorem 1 ensures that (UK ∗ )S is a stable frame operator, and the
maximum ω given known set S, as well as interpolation in case of solution can be found by computing the pseudo inverse of (UK ∗ )S .
non-bandlmited signals. Thus, the interpolated graph signal on the unknown set S c is given
by
3. PROPOSED INTERPOLATION
−1
g∗ (S c ) = (UK ∗ )S c ((UK ∗ )S )t (UK ∗ )S ((UK ∗ )S )t f (S).
(7)
To make a practical use of Pesenson’s result in Theorem 1, we first Finally, the interpolated signal is computed as: f̂ = D−1/2 ĝ.
present a result to compute the maximum ω, such that any signal in Note that, while other global methods such as [9, 10] also pro-
P Wω (G) can be reconstructed, given a subset of known samples in pose similar least-square reconstruction solutions, the choice of
S on any arbitrary graph G. number of eigenvectors K in these methods is heuristic. In our
proposed method K ∗ is chosen specifically to be the number of
Proposition 2 Given a graph G with normalized Laplacian matrix eigenvalues below the cut-off frequency ωS∗ given in Proposition 2,
L, known set S and unknown set S c , let (L2 )S c be the submatrix of which depends on the known set of nodes, and the topology of the
L2 containing only the rows and columns corresponding to unknown graph. The optimality of choosing ωS∗ as cut-off frequency can be
set S c . Then the known set S is a uniqueness set for all signals f ∈ justified as follows:
P WωS (G) with ωS∗ = σmin , where σmin 2
is the smallest singular Let ωS is the chosen cut-off frequency. If signal f ∈ P WωS (G),
2
value of (L )S c . then the proposed reconstruction is perfect (loss-less), hence opti-
mal. For f ∈ / P WωS (G), our proposed method still provides a sta-
Proof Referring to Definition 2, let φ be a signal in L2 (S c ), i.e., ble least-square solution. The solutions with ωS < ωS∗ are clearly
φ = [0(S)t φ(S c )t ]t . We have suboptimal in this case since the solution space of ωS is contained
in the solution space of ωS∗ . For ωS > ωS∗ , the reconstruction may
kLφk2 < φ(S c ) (L2 )S c φ(S c ) > sometimes produce less error, but it is not guaranteed to be stable
= = R(L2S c ) (4) (i.e., matrix (UK ∗ )S might not be a frame). In Section 4, we show
kφk 2 kφ(S c )k2
that choosing ωS > ωS∗ leads to poorer results.
where R(.) is the Rayleigh quotient of a matrix. It can be shown that
R((L2 )S c ) is always greater than the minimum eigenvalue σmin
2
of 3.2. Bilateral Link-weight Adjustment
the matrix (L2 )S c . Thus,
Unlike regular signals, the smoothness of a graph signal depends
2 both on the signal values and the underlying graph. This leads to
kLφk 2 1
≥ σmin ⇒ kφk ≤ kLφk (5) the question of whether we can modify the graph to adapt to the
kφk2 σmin given signal so that the signal is more band-limited on the simpli-
fied graph, thus leading to less interpolation error. The simplifica-
Thus, S c is a Λ-set with Λ = σmin1
. By Theorem 1, S is a unique- tion makes sense in many cases such as in recommendation systems,
ness set for all signals f ∈ P Wω (G) with ωS∗ = σmin . where the underlying graph is the result of observing average corre-
lation over a set of training users (multiple instances), and the signal
Note that, ωS∗ computed above is the maximum possible value that corresponds to a single test user. We take inspiration from image
satisfies the sufficient conditions in Theorem 1. We term ωS∗ as the processing where this kind of signal adaptive filtering is achieved by
cut-off frequency for reconstruction, since any graph-signal below bilateral filters [15].
this frequency can be perfectly reconstructed from its DU signal on In our proposed method we use g = D1/2 f as the signal to
S. be interpolated. From (1), we observe that g can be made more
bandlimited by minimizing ||Lg||. Define an error function:
3.1. Interpolation method ζ = Lg = (I − W)g = D1/2 (I − D−1 W)f . (8)
Given the cut-off frequency ωS∗ computed from Proposition 2, the Clearly, minimizing ζ at each node minimizes ||Lg||. The value of
reconstruction of any graph signal is done by least-square projection ζ at node i can be written as:
of the corresponding DU signal onto the P Wω (G) space. Let K ∗ !
be the number of eigenvalues of the Laplacian matrix L, less than √ 1 X
ζ(i) = di f (i) − wij f (j) , (9)
ωS∗ . Define UK ∗ as the matrix containing first K ∗ eigenvectors, and di j
(UK ∗ )S as the submatrix of UK ∗ containing rows corresponding
to set S. The first eigenvector of a connected graph is u1 = D1/2 1, which is proportional to the difference of f (i) with the weighted av-
which is not constant for graphs with irregular degrees. Therefore, erage (the weight being link-weights) of nodes directly connected to

5447
node i. Thus, by adapting the weights wij to be inversely propor- with or without bilateral weighting are very close to PMF method,
tional to the absolute difference |(f (i) − f (j)| at every node, we can with the interpolation with bilateral weights performing slightly bet-
minimize the error ζ, and hence ||Lg||. Following similar intuition ter. We also observed that choosing K = K ∗ + 10 in LS method
as in bilateral filters, we modify the weights between nodes in S as: leads to poorer results. The Figure 2(b) shows another RMSE plot
where users are grouped by the number of training samples, with
|f (i) − f (j)|2 x-axis showing those groups. We observe that both the LS method
ŵij = wij .exp(− ) ∀{i, j ∈ S}, (10)
θ2 and kNN method perform significantly worse when the number of
available training samples are small. The effect of applying bilateral
where parameter θ is chosen to be the mean rating over all training
weighting in our proposed methods is also most visible here.
samples. Note that in (10), we can only change the weights between
The PMF method predicts the ratings of all movies for all users
two known nodes. Our proposed interpolation algorithm is given in
simultaneously by factorizing the whole N × M rating matrix. It is
Algorithm 1. The complexity of our method is O(K ∗ |V|2 ), primar-
based on an iterative update rule and requires O(N M P ) operations
ily due to the least-square projection steps (5 − 7) in Algorithm 1.
per iteration where P is the size of the latent space. Theoretically,
any change to the rating matrix would require the PMF system to be
Algorithm 1 Proposed Graph Interpolation Method retrained on all users. However, in our method, the process of com-
Require: G0 = (V, E): Initial Graph, f : DU signal puting the movie graph is decoupled from the process of predicting
1: Compute normalized Laplacian matrix L. ratings for a given user. Accommodating a few new ratings into the
2: Compute ωS ∗ as the square-root of the smallest eigenvalue of (L2 ) c .
S systems is fast as it only affects a local portion of the movie graph.
3: Modify known nodes’ link-weights as bilateral weights using (10). Once the movie graph is fixed, the proposed method allows us to
4: Recompute normalized Laplacian matrix L. predict the ratings of movies for each user separately in O(K ∗ M 2 )
5: Compute K ∗ eigenvectors of L corresponding to eigenvalues λ < ωS ∗.
operations. Thus, assuming K ∗ ≈ P , the proposed method is faster
6: Compute g = D1/2 f .
7: Compute ĝ: (ĝ)S = (g)S and (ĝ)S c using (7).
(i.e., M 2 < N M ) than PMF, when ratings of items change fre-
8: Interpolated signal f̂ = D−1/2 ĝ.
quently and the recommendations need to be calculated only when a
user request them. Further , it may be possible to reduce complexity
in our method by using simplified filtering operations.
4. EXPERIMENTS
3
LS projection (K= K* + 10)
We apply proposed interpolation method for collaborative filtering kNN (k = 30)
PMF
in recommendation systems. The input in this problem is a partially
Cumulative RMSE

2.5
Proposed method (K = K*)
Proposed Method with bilateral weights
observed user-item rating matrix R, such that R(u, m) is the rating
2
given by user u to the movie m. Based on this information, the
system predicts new user-movie ratings. For empirical evaluation, 1.5
we choose the MovieLens 100k [16] dataset containing 100k user-
movie-rating triplets from N = 943 users and M = 1682 movies. 1

The ratings are integer values between 1 and 5. We use the 5-fold 0 100 200 300 400
Number of training samples
500 600 700

cross validation data available in [16], which consists of 5 disjoint

random sets of 20k triplets. At each iteration, one of these sets is
used for testing and remaining sets for training.
We first compute a movie graph G0 from the training samples,
by computing cosine similarity [17] between every pair of movies.
For each test user u, we define S to be the set of movies with known
ratings, and and U to be the set of test movies. We compute the
subgraph Gu = (S ∪ U, Eu ), corresponding to subset S ∪ U of
nodes. We define DU signal fu for u to be of size |U ∪ S|, with
fu (U) = 0 and fu (S) equal to known ratings. Subsequently, we
compute interpolated signal f̂u by using proposed method given in
Algorithm 1, with Gu and fu as inputs. In our preliminary analysis, Fig. 2: RMSE of different prediction algorithms with the number of training
we compare the performance of our proposed method with two most samples on MovieLens dataset .
popular algorithms for collaborative filtering: 1) kNN method [17]
(with k = 30), and 2) probabilistic matrix factorization (PMF) [18]
with 10 latent features. Our proposed method is similar to the least
square (LS) interpolation of [9], except that we operate based on nor- 5. CONCLUSIONS AND FUTURE WORK
malized Laplacian matrix, and choose K ∗ specified by ωS∗ . To show
that this K is a good choice, we also implement the method in [9] We proposed a novel method for interpolation of graph structured
with K = K ∗ + 10 (i.e., with 10 additional eigenvectors), respec- data using signal processing tools. We formulated the interpolation
tively. We use RMSE between predicted values and actual values to problem as reconstruction of a graph-signal from its DU sampled
measure performance. The predicted values less than 0 (more than signal. The success of this method in preliminary experiments opens
5) are set to 0 (5) before computing RMSE. Figure 2 plots the RMSE up many new opportunities and challenges. The future work includes
of prediction as a function of number of training samples available. improving interpolation accuracy by a) finding out both necessary
Figure 2(a), shows cumulative RMSE of various implemented meth- and sufficient conditions for reconstruction (Theorem 1 is only a
ods. Observe that kNN method perform the worst of all method, sufficient condition), and b) understanding and formulating bilateral
and PMF method performs the best. Our proposed methods both weighting step as an optimization problem.

5448
 6. REFERENCES

[1] M. Weber and S. Kube, “Robust Perron cluster analysis for var-
ious applications in computational life science,” in CompLife,
2005, pp. 57–66.
[2] Z. Huang, W. Chung, T. Ong, and H. Chen, “A graph-based
recommender system for digital library,” pp. 65–73, 2002.
[3] M. Girvan and M. E. Newman, “Community structure in social
and biological networks.,” Proc Natl Acad Sci U S A, June
2002.
[4] D. I Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Van-
dergheynst, “Signal processing on graphs: Extending high-
dimensional data analysis to networks and other irregular data
domains,” to appear SPM 2012, also in arXiv’12, , no.
arXiv:1211.0053, Dec. 2012.
[5] R.I. Kondor and J. Lafferty, “Diffusion kernels on graphs and
other discrete structures,” in Proc. ICML, 2002, pp. 315–322.
[6] S. Hoche, P. Flach, and D. Hardcastle, “A fast method for
property prediction in graph-structured data from positive and
unlabelled examples,” in ECAI 2008, 2008, pp. 162–166.
[7] J. Bennett, S. Lanning, and N. Netflix, “The netflix prize,” in
In KDD Cup and Workshop in conjunction with KDD, 2007.
[8] J. Chen, H. Fang, and Y. Saad, “Fast approximate KNN graph
construction for high dimensional data via recursive lanczos
bisection,” J. Mach. Learn. Res., vol. 10, pp. 1989–2012, dec
2009.
[9] M. Belkin and P. Niyogi, “Semi-supervised learning on rie-
mannian manifolds,” pp. 209–239, 2004.
[10] L. Grady and E. L. Schwartz, “Anisotropic interpolation on
graphs: The combinatorial dirichlet problem,” Tech. Rep.,
Boston University, 2003.
[11] I. Pesenson, “Sampling in Paley-Wiener spaces on combina-
torial graphs,” Trans. Amer. Math. Soc, vol. 360, no. 10, pp.
5603–5627, 2008.
[12] S.K. Narang and A. Ortega, “Downsampling graphs using
spectral theory,” in ICASSP ’11., May 2011.
[13] E. B. Davies, G. M. L. Gladwell, J. Leydold, and P. F. Stadler,
“Discrete nodal domain theorems,” Linear Algebra and its Ap-
plications, vol. 336, no. 1-3, pp. 51 – 60, 2001.
[14] S.K. Narang and Ortega A., “Perfect reconstruction two-
channel wavelet filter-banks for graph structured data,” IEEE
trans. on Sig. Proc. 2012, vol. 60, no. 6, June 2012.
[15] C. Tomasi and R. Manduchi, “Bilateral filtering for gray and
color images,” in Sixth International Conference on Computer
Vision, Jan 1998, pp. 839–846.
[16] “MovieLens dataset,” as of 2003, http://www.
grouplens.org/.
[17] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Item-based
collaborative filtering recommendation algorithms,” in WWW
’01, 2001, pp. 285–295.
[18] Ruslan Salakhutdinov and Andriy Mnih, “Probabilistic matrix
factorization,” in Advances in Neural Information Processing
Systems, 2008, vol. 20.

5449