# Collaborate with strangers to find own preferences.

**ABSTRACT** Abstract We consider a model with n players and m objects. Each player has an unknown,grade for each object, modeled by a “preference vector” of length m. A player can learn his grade for an object by probing that object, but performing a probe incurs cost. The goal of the players is to learn their own evaluations of objects with minimal cost, by adopting the results of probes performed by other players. To facilitate communication, we assume that players collaborate by posting their grades for objects on a shared billboard: reading from the billboard is free. We consider players whose preference vectors are popular, i.e., players whose preferences are common to many other players. We present distributed and sequential algorithms to solve the problem with logarithmic cost overhead. Submitted as a regular presentation. Please consider as a brief announcement as well.

**0**Bookmarks

**·**

**112**Views

- [Show abstract] [Hide abstract]

**ABSTRACT:**We consider a model of recommendation systems, where each member from a given set of players has a binary preference to each element in a given set of objects: in- tuitively, each player either likes or dislikes each object. However, the players do not know their preferences. To nd his preference of an object, a player may probe it, but each probe incurs unit cost. The goal of the play- ers is to learn their complete preference vector (approx- imately) while incurring minimal cost. This is possible if many players have similar preference vectors: such a set of players with similar \taste" may split the cost of probing all objects among them, and share the results of their probes by posting them on a public billboard. The problem is that players do not know a priori whose taste is close to theirs. In this paper we present a distributed randomized peer-to-peer algorithm in which each player outputs a vector which is close to the best possible ap- - SourceAvailable from: eng.tau.ac.il
##### Article: Asynchronous recommendation systems

[Show abstract] [Hide abstract]

**ABSTRACT:**We consider the following abstraction of recommendation systems. There are n players and m objects, and each player has an arbitrary binary preference grade ("likes" or "dislikes") for each object. The problem is that these preferences are not known, and the goal of the players is to discover their own preferences. To do that, a player can probe each object, thereby directly finding his preference grade for the objects. However, probing an object incurs cost. To save on cost, players post the results of their probes on a public billboard: writing and reading from the billboard is free. The idea is that cost can be reduced if players with similar preferences share the load of probing, but such similarities are not a priori known to the players. In a synchronous recommendation system, players probe in global rounds, and in an asynchronous system, players probe in an order determined by an arbitrary schedule. In this paper we present thefirst asynchronous recommendation systems that can reconstruct the preferences of players under adversarial asynchronous scheduling, with polylogarithmic overhead in cost with respect to the best possible. We present algorithms both for exact and approximate preference reconstructions.01/2007; - SourceAvailable from: Keno Albrecht[Show abstract] [Hide abstract]

**ABSTRACT:**In this paper, we present a lightweight but powerful plug-in container which provides advanced features such as dynamic class loading, dependency, configu-ration, and security management. We highlight the deployment mechanism which allows to publish, in-stall, and update plug-ins from arbitrary sources at runtime. Furthermore, we introduce the Trooth voting and trust system which is used to assess plug-ins but has been designed as a general rating mechanism. Finally, we present the extensible ar-chitecture of the Spamato spam filter framework and show how it employs the plug-in mechanism and the Trooth system in a real-world applica-tion.09/2006;

Page 1

Collaborate With Strangers to Find Own Preferences∗

Extended Abstract

Baruch Awerbuch†

Yossi Azar‡

Zvi Lotker§

Boaz Patt-Shamir¶

Mark R. Tuttle?

Abstract

We consider a model with n players and m objects. Each

player has a “preference vector” of length m that models

his grade for each object. The grades are unknown to the

players. A player can learn his grade for an object by

probing that object, but performing a probe incurs cost.

The goal of a player is to learn his preference vector with

minimal cost, by adopting the results of probes performed

by other players. To facilitate communication, we assume

that players collaborate by posting their grades for objects

on a shared billboard: reading from the billboard is free.

We consider players whose preference vectors are popular,

i.e., players whose preferences are common to many other

players. We present distributed and sequential algorithms

to solve the problem with logarithmic cost overhead.

Categories and Subject Descriptors

C.2.4 [Computer-Communication Networks]: Dis-

tributed Systems; F.2.2 [Analysis of Algorithms and

∗Research partly done in HP Cambridge Research Lab (CRL).

†Dept. of Computer Science, Johns Hopkins University.

Email: baruch@cs.jhu.edu. Supported by NSF grant ANIR-

0240551 and NSF grant CCR-0311795.

‡School of Computer Science, Tel Aviv University, Tel Aviv

69978, Israel. Email: azar@tau.ac.il.

in part by the German-Israeli Foundation and by the Israel

Science Foundation.

§Kruislaan 413 P.O. Box 94079, CWI, 1090 GB Amsterdam,

The Netherlands. Email: lotker@cwi.nl.

¶Dept. of Electrical Engineering, Tel Aviv University, Tel Aviv

69978, Israel. Email: boaz@eng.tau.ac.il.

ported in part by Israel Ministry of Science and Technology.

?HP Cambridge Research Lab, One Cambridge Center, Cam-

bridge MA 02142, USA. Email: mark.tuttle@hp.com.

Research supported

Research sup-

Permission to make digital or hard copies of all or part of this work for

personal or classroom use is granted without fee provided that copies are

not made or distributed for profit or commercial advantage and that copies

bear this notice and the full citation on the first page. To copy otherwise, to

republish, to post on servers or to redistribute to lists, requires prior specific

permission and/or a fee.

SPAA’05, July 18–20, 2005, Las Vegas, Nevada, USA.

Copyright 2005 ACM 1-58113-986-1/05/0007 ...$5.00.

Problem Complexity]: Nonnumerical Algorithms and

Problems

General Terms

Algorithms, Theory.

Keywords

recommendation systems, collaborative filtering, random-

ized algorithms, electronic commerce, probes, billboard.

1.Introduction

Most people encounter recommendation systems on a

daily basis in many situations, such as buying a book,

renting a movie, choosing a restaurant, looking for a ser-

vice provider etc. Recommendation systems are supposed

to help us make choices among the different alternatives

that are available to us. The introduction of the Inter-

net had made recommendation systems even more criti-

cal than ever, since entities (such as users and products)

on the Internet are typically virtual, and may be located

on the other side of the globe. As a result, a user may

have very little side-information to work with, and must

rely on computerized recommendation systems. The idea

in these systems is to take advantage of the existence of

many users with similar opinions: a set of users with sim-

ilar preferences (called type) can collaborate by sharing

the load of searching the object space and identifying ob-

jects they deem as good. However, players belonging to

the same type do not know of each other a priori, and

therefore, implicitly or explicitly, significant effort must

be directed into identifying potential collaborators (while

avoiding possible crooks).

Much of the extant research has concentrated on try-

ing to find a good object for each user; in this paper we

address a more general question, where the goal of the

users is to try to find what are their opinions on all ob-

jects. The advantage of having such an algorithm is that

it allows users to identify others who share their opinions,

which may be useful in the long range: When a new batch

of objects arrives, much effort can be saved by using the

opinions of users who have already been identified as be-

longing to my own type. Moreover, as we show in this

Page 2

paper, the cost of finding the complete set of grades (and

hence identifying users with similar preferences) is com-

parable with the cost of finding just one good object for

each user.

Roughly speaking, our model is as follows (formal def-

initions are provided in Section 2). We assume that there

are n users and m objects, and that each user has an

unknown grade for each object, i.e., each user can be rep-

resented by his preference vector. The goal is for each user

to find his complete preference vector.1The problem is

that players do not know what is their preference for an

object unless they probe it, but probing is a costly opera-

tion (e.g., the user needs to buy the product). Neverthe-

less, players would like to predict what is their preference

for items they did not probe. To facilitate this, the system

provides a shared “billboard,” on which players can post

the results of their probes, thus allowing players with sim-

ilar preferences to split the probes among them and share

the results. The existence of a billboard does not immedi-

ately solve the problem: the question now is which grades

should one adopt, since players may have arbitrarily dif-

ferent preference vectors. To allow a non-trivial solution

to exist, we assume that some preference vectors are pop-

ular, namely many players have them. Such players are

said to belong to the same type. Ideally, players of the

same type should share the work and the results among

them; however, it is not known who belongs to which type

when the algorithm starts.

Symmetric operation.

imposed on any reasonable solution is that it must be

symmetric, i.e., with high probability, all players execute

more-or-less the same amount of work.

tion is lifted, it is easy to come up with simple asym-

metric “committee-style” solutions to the problem (see

Sec. 5.1). However, such algorithms, while possibly suit-

able for centralized settings, seem unrealistic in the dis-

tributed model, for the following reasons (see [5, 2] for

more discussion of the issue):

One important requirement

If this restric-

• Lack of fairness: the number of objects may be

huge, and the tolerance of players for doing a lot

of work for others may be in short supply.

• Distrust of committees: small-size committees are

prone to easy manipulation by adversaries (bribery,

intimidation, other forms of corruption) and thus

cannot be trusted to be objective judges on impor-

tant matters.

1.1 Lower bounds

Before we state our results precisely, let us sketch two

trivial lower bounds on the cost. Consider first a nearly

ideal scenario, where all players in a given type T know the

identities of all other type-T players, but they don’t know

what is the type-T preference vector. Even if the players

coordinate perfectly to minimize their overall cost to find

the vector, each object must be probed at least once by

some type-T player. Therefore, a lower bound of Ω(m) on

1We arbitrarily write “he” for users, but it should be read as

“he or she.”

the total cost to the players of type T is unavoidable. Now

consider a second nearly ideal scenario, in which all type

vectors are common knowledge, and the only question a

player must answer is what type does he belong to. It

is not hard to see that if there are k possible types, Ω(k)

expected probes by each player are unavoidable. It follows

that the total cost of all players of a single type of size

Ω(n/k) is Ω(n). Combining the lower bounds from the

two scenarios above we conclude that no algorithm can

solve the problem is o(m + n) cost over all players of a

given type.

1.2Our results

In this paper we take another step toward understand-

ing recommendation systems. Informally, we present al-

gorithms (in the simplified model sketched above), which

demonstrate that if one’s unknown preferences are known

not to be esoteric (i.e., it is known that there are quite

a few other players with the same unknown preferences),

then one can find his own preferences without performing

more than a logarithmic number of probes. More specifi-

cally, we present symmetric algorithms that approach the

above lower bound to within a logarithmic factor. The

algorithms are randomized, and they are guaranteed to

work with probability 1 − n−Ω(1). The algorithms rely

on knowing a lower bound on the popularity of the type

in question. Formally, the algorithms use a parameter

0 < α ≤ 1 such that there are at least αn players with

the preference vector we would like to output.

Our main result is a parallel algorithm called D, whose

running time is O

αn

tal number of probes done by users of any given type is

O ((n + m)logn). Our algorithm works for any set of user

preferences: only the running time depends on the type

size.

We also present a sequential algorithm which slightly

improves the total number of probes for large α, while

maintaining the crucial fairness property that all users

make approximately the same number of probes. The se-

quential algorithm (called S1) ensures that the total num-

ber of probes done collectively by players in any given type

is at most O

α

time of the latter algorithm may be high.

Detailed comparison with previous results is given in

Sec. 5, after the presentation of our algorithms. As a one

line summary, let us emphasize at this point that unlike

our algorithms, all previous solutions to this problem are

either asymmetric (i.e., some players do linear work), or

cannot deal with all inputs.

?log n

?m

??. This means that the to-

?mlogn +n

?. Unfortunately, the running

Paper organization. The remainder of this paper is

organized as follows. In Section 2 we formalize the model.

In Section 3 we present our sequential algorithm, S1. In

Section 4 we present our parallel algorithm, D.

concluding remarks and open problems are presented in

Section 6. In Section 5 we describe existing work. In

section 6 we conclude with some open problems.

Some

Page 3

2.Model

Our model is formally defined as follows. There are

n players and m objects. For each object i ∈ {1,...,m}

and player j ∈ {1,...,n}, there exists a value vj

called the grade or value of object i to player j. The

vector vj

= (vj

player j. A type is a vector v ∈ {0,1}m. A player j is

said to belong to type v if his preference vector vjis equal

to v. We say that the popularity of a type is α, for some

0 ≤ α ≤ 1, if there are at least αn players of that type.

An execution is a sequence of steps, where in each step

some players move. A basic move of a player j is to reveal

his value vj

player j probed object i, and say that vj

i. Each player may probe one object in a step, but many

players may probe the same object simultaneously. The

players communicate by means of a shared billboard. To

model its existence, we assume that the results of all past

probes are freely available to all players.

An algorithm in our model is a function that decides

which player probes which object, based on past probe

results, and possibly using a tape of random bits. Al-

gorithms are parametrized by the number of players and

objects (called n and m, respectively), a confidence para-

meter (called c), and a lower bound α on the popularity of

the type whose vector the algorithm needs to output. The

individual or aggregate cost of an algorithm is the number

of probes incurred by either an individual player or by all

players of a given type, respectively. The running time of

an algorithm is the total number of parallel steps taken

since the start of the algorithm until all players (possibly

of a given type) have terminated.

By the end of an execution of an algorithm, each

player outputs a vector in {0,1}m. An algorithm is said to

be correct with probability p for players of type T if with

probability p, the output vector of each player of type

T is exactly the player’s preference vector. Probability

is taken over the tape of random bits. The algorithms

presented in this paper are correct with high probability,

meaning that the probability of failure is n−Ω(c), where c

is the confidence parameter of the algorithm.

We stress that the only way randomization is used by

algorithms in this paper is in assuming that the identities

of players and the identities of objects form random per-

mutations of {1,...,n} and of {1,...,m}, respectively.

i∈ {0,1},

def

1,vj

2,...,vj

m) is the preference vector of

ifor some object i. In this case we say that

iis the vote of j on

3. Sequential Algorithm

In this section we present a simple algorithm called S1

with low cost, but high running time. We fix our attention

on a single type of popularity α > 0. Algorithm S1 works

for any value of 0 < α ≤ 1. Its cost is better than the

cost of Algorithm D (presented in Section 4) for α =

Ω(1/logn).

The algorithm works as follows (see Figure 1). The al-

gorithm proceeds by determining the value of objects one

at a time. This is done by letting random players probe

the object until the players can conclude, with high prob-

ability, what is the value of the object (Step 3). The high

probability test is based on the lower bound on the num-

ber of players with the same type: if there are too few

players who think that the value of the object is v, then,

given the lower bound on the popularity of the type, the

value of that object to that type is not v . The thresh-

old is computed in Step 2. It may be the case, however,

that the high probability test will never end with a single

value for the object (say when α < 1/2). In this case all

players might end up testing the object. Following this

strategy naively will lead to the trivial cost of nm probes.

To avoid that, we use the following simple idea. Each

player locally marks each other player as either “quali-

fied” or “unqualified.” The interpretation of the marks is

that a player j?is marked “unqualified” by player j only

if there is evidence that j?does not belong to the type of

j. Initially, all players are marked “qualified” by every-

one. The marks are updated by subroutine call: Once

an object value is determined, all players who voted for

another value are “disqualified” by the current player i.e.,

that player disregards them from that point and onward.

The important property of the algorithm is that all

players of the same type maintain a consistent view of

which players are qualified and which aren’t.

Lemma 3.1. For any object i, all players of type T ex-

ecute call(i,v) with the same value v. Furthermore, they

all have the same set of qualified players.

Proof: By induction on the objects. For the basis of the

induction we have that all players of type T have the same

set of players as qualified when the algorithm starts, since

all players are qualified. For the induction step, consider

the point when an object is called. This happens either

in Step 3 or in Step 4b. In the former case, all players of

type T will call the same value because they do it based on

the same billboard and, by induction, they use the same

set of qualified players. In case an object is called in Step

4b, all players of type T have the same local value for the

object by definition of type. The agreement on the set of

qualified players follows in both cases.

Lemma 3.1, in conjunction with the following lemma,

implies correctness of the output.

Lemma 3.2. If a player j of a type whose popularity is

at least α executes call(i,v), then P

?vj

i= v

?< n−c.

Proof: If player j executes call(i,v) at Step 4b, then j

actually probed i, and vj

call(i,v) at Step 3, then we have that only the value v

received more than θi of the qi qualified votes. In this

case it is straightforward to bound the probability that

vj

versions of [10]). The voters are chosen at random (Step

4a), and therefore, by assumption on the popularity of the

type, the probability that a random vote is vj

α. Hence

i= v with certainty. If j executes

i= v using the Chernoff bound as follows (we use the

iis at least

P

?

vj

i= v

?

≤

<

<

P

?

vj

ireceived less than θi votes

?

e−αqi/8

e−c ln n= n−c.

The first inequality above follows from Step 3; the sec-

ond inequality follows from Chernoff, since the expected

Page 4

Algorithm S1:

For each object i in turn do:

1. Let the number of votes of qualified players for object i be qi.

2. Let θi ←α

3. If there is only one value v with more than θi qualified votes, then execute call(i,v), and proceed to the next

object.

2(qi− 8clnn).(θi is a threshold)

4. Otherwise(multiple values pass threshold)

(a) If not all players have probed i, let a random player probe object i and go to Step 1.

(b) Otherwise (all players have already probed i), let v be the value seen by the local player. Execute call(i,v)

and proceed to the next object.

Subroutine call(i,v):

1. Output the value v for object i.

2. Mark all players whose vote on i is different than v as “unqualified.”

Figure 1: Pseudo-code for Algorithm S1.

number of votes for vj

inequality follows from the fact that when call(i,v) is

executed, θi is strictly positive, and hence, by definition

of θi, we αqi > 8clnn at this point.

iis at least αqi > 2θi; and the last

We now bound the cost of S1. To do that, we count

the number of qualified votes, i.e., votes of players that

are considered qualified when Step 1 is executed (note

that some of the players who cast such votes may later be

disqualified).

Lemma 3.3. The

throughout the execution is at most O

total number ofqualifiedvotes

?mlnn +1−α

i denote the

αn

?.

Proof Sketch:

total number of qualified players who vote on i when

call(i,v) is executed. Obviously, the total number of

probes of type-T players is at most

iq?

one hand, each player may be disqualified at most once,

and hence the total number of disqualifications, through-

out the execution, is at most (1 − α)n.

hand, note that whenever call(i,v) is executed, at least

α

2((q?

i was not called in the previous round, when there were

q?

α

2(q?

than v. It follows that

For each object i, let q?

?

iq?

i. We bound

?

iby counting the number of disqualified players. On

On the other

i−1)−8clnn) players are disqualified: this is because

i− 1 qualified votes, and therefore there were at least

i−1 − 8clnn) players who voted for a value different

(1 − α)n ≥

m

?

i=1

α

2

?q?

i− 1 − 8clnn

?

,

i.e.,

m

?

i=1

q?

i ≤ O

?

mlnn +1 − α

α

n

?

.

We summarize with the following theorem.

Theorem 3.4. Suppose that there are at least αn play-

ers from a certain type, and that they all run Algorithm

S1. Then with probability 1 − n−Ω(1)they will all output

the correct vector, after executing O(mlogn +n

α) probes.

Since in Algorithm S1 only one player probes in each

round, its time complexity is equal to the number of

probes. It is easy to speed up the algorithm somewhat.

For example, we can have all players probe all objects

in parallel until there are O(log n) samples per object,

and then apply the algorithm only to the undetermined

objects. We do not elaborate any further on this idea,

since in the following section we present an algorithm with

nearly the best possible parallel time.

4. Parallel Algorithm

In this section we present our main result, a recur-

sive parallel algorithm called D. This algorithm achieves

near-optimal time complexity while maintaining the cost

modest.

The idea in Algorithm D is as follows (see Figure 2).

Given a set of players and a set of objects, we split the

players and objects into two subsets each, let each half

of the players recursively determine the values of half of

the objects, and then merge the results. Figure 3 gives a

visual representation of the situation after returning from

the recursive call. Merging results means filling in the

missing entries. Merging is done as follows. First, each

player finds the preference vectors that are sufficiently

popular in the other half (Step 4). This leaves only a

few candidate vectors the players needs to choose from

to complete his row. Testing objects with disputed val-

ues (Steps 5–6), the player eliminates at least one vector

in each probe, and eventually, the player adopts the last

surviving vector.

In the algorithm, we assume that the partition of

sets of players and objects into halves is done by a fixed

function. For example, if the current set of players is

P = {j,j + 1,...,j + k − 1} and the current set of ob-

jects is O = {i,i + 1,...,i + k?− 1}, then P?may be

Page 5

Algorithm D(P,O):

1. If min(|P|,|O|) <8c ln n

2. (Otherwise) Let P?be the half of P such that j ∈ P?, and let P??be the other half. Let O?be the half corresponding

to P?in O, let O??be the other half of O.

3. Execute D(P?,O?).

(upon returning, values for all objects in O?were output by all players in P?, and values for all objects in O??were

output by all players in P??)

4. Scan the billboard. Let V be a set of vectors for O??such that the each vector in V is voted for by at least α/2 of

the players in P??.

5. Let C ⊆ O??be the set of objects for which there are different votes in V .

(P is a set of players and O is a set of objects)

α

then probe all objects in O, output their values, and return.

(The partition of P and O is a fixed function of the IDs)

6. While C = ∅ do

(a) Choose an arbitrary object i ∈ C. Probe i, and let the probe result be denoted by v.

(b) Remove from V vectors whose value on i is not v.

(c) Update C to contain all objects on which surviving vectors in V differ.

7. Output the (unanimous) votes of surviving vectors in V for all objects in O??and return.

Figure 2: Pseudo-code for Algorithm D as executed by player j. When The algorithm returns, the players in P have

output values for all objects in O.

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

v1

v2

...

vn?

1

1

v1

v2

...

vn?

2

2

...

...

...

...

v1

v2

m?

m?

...

12

vn?

m?

vn?+1

m?+1

vn?+2

m?+1

...

v2n?

m?+1

vn?+1

m?+2

vn?+2

m?+2

...

v2n?

m?+2

...

...

...

...

vn?+1

2m?

vn?+2

2m?

...

v2n?

2m?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

Figure 3: Matrix representation of the results after Step

3 in Algorithm D. Each row represents a player, and each

column represents an object. P?is the top half of the rows,

and O?is the left half of the columns. |P?| = |P??| = n?

and |O?| = |O??| = m?. The entries in two quadrants of

the matrix are unknown.

defined by P?= {j,j + 1,...,j + k/2 − 1} and similarly

O?= {i,i + 1,...,i + k?/2 − 1}. The sets P??and O??are

defined to be the remainders P \ P?and O \ O??, respec-

tively. Since we assume that the identities of players and

objects form a random permutation, these choices define

random partitions of P and O into two halves each (we

assume, for convenience, that n and m are powers of 2).

In the analysis, we focus our attention on a certain

type T with popularity α.

Lemma 4.1. Fix an invocation of D(P,O) with |P| ≥

8c ln n

α

. Then with probability at least 1−n−c, there are at

least α|P|/2 type-T players in P in that invocation.

Proof: The lemma follows from the fact that P is a ran-

dom subset of the players, and hence the expected number

of type-T players in P is α|P|. Using Chernoff we get that

P

less than

2

type-T players in P

?

α|P|

?

<eα|P|/8

≤

n−c,

since |P| ≥8c ln n

α

.

Lemma 4.2. The total number of probes for a single

player in the execution of the algorithm is O

?log n

α

?m

n

??.

Proof Sketch: Step 1 is taken only once by a player.

If Step 1 is taken because |P| <

player set and the object set are halved in each recursive

call, we may conclude that when Step 1 is taken, |O| <

m·8c ln n

In any case, Step 1 incurs at most O(m log n

O(

n

The only other probes made by the algorithm are

made in Step 6a. In any given invocation of the algo-

rithm, each player makes at most |V | probes at Step 6a.

Since by Step 4 each member of V represents at least an

α/2 fraction of the players, we have that |V | ≤

the depth of the recursion is log

have that the total number of probes done in Step 6a by

each player, throughout the execution of the algorithm, is

O

α

We summarize in the following theorem.

8c ln n

α

, then since the

αn. Otherwise, Step 1 is taken because |O| <8c lnn

α

.

αn

+

log n

α) =

?m

?log n

α) probes.

2

α. Since

n

8c ln n/α= O(logn), we

?log n

?. The result follows.

Theorem 4.3. Suppose that there are Θ(αn) players

from a certain type T, and that they all run Algorithm D.

Then with probability 1 − n−Ω(1)they will all output the

correct vector, after executing O((m + n)logn) probes in

O

αn

We note that both the total cost and the individual

running time of Algorithm D are the best possible, up to

a factor of O(logn).

?log n

?m

??rounds.

Page 6

5.Related work

Most prior research on recommendation systems fo-

cused on a centralized, off-line version of the single rec-

ommendation problem, where the algorithm is presented

with a lot of historical preference data, and the task is

to generate a single recommendation that maximizes the

utility to the user. This is usually done by heuristically

identifying clusters of users [12] (or products [13]) in the

data set, and using past grades by users in a cluster to pre-

dict future grades by other users in the same cluster. Sin-

gular Value Decomposition (SVD) was also shown to be

an effective algebraic technique for the off-line single rec-

ommendation problem [14]. Some of these systems enjoy

industrial success, but they are known to perform poorly

when prior data is less than plentiful [15], and they are

quite vulnerable even to mild attacks [9, 11].

Algorithmic results. Theoretical studies of recommen-

dation systems usually take the latent variable model ap-

proach: a stochastic process is assumed to generate noisy

observations, and the goal of an algorithm is to approxi-

mate some unknown parameters of the model. Kumar et

al. [8] study the off-line problem for a model where pref-

erences are identified with past choices (purchases). In

this model there are clusters of products. Each user has a

probability distribution over clusters; a user first chooses

a cluster by his distribution, and then chooses a product

uniformly at random from that cluster. The goal is to

recommended a product from the user’s most preferred

cluster. Kleinberg and Sandler [7] generalized this model

to the case where the choice within a cluster is governed

by an arbitrary probability distribution, and also consider

the mixture model, in which each cluster is a probability

distribution over all products.

Singular value decomposition approaches. Azar et

al. introduced in [3] the idea of using SVD to reconstruct

the unknown preference vectors. This idea was later used

by Drineas et al. [5]. The attractive feature of the SVD

method is its ability to deal with some noise. However,

SVD is inherently incapable of handling preference vec-

tors that cannot be approximated by a low rank matrix.

By contrast, our algorithm requires only a bound on the

individual type size: what happens with other types is

completely irrelevant. As a result, our algorithms can deal

with any input, and the running time for a given type is

nearly optimal, while the correctness of SVD-based al-

gorithms is dependent on players from other types. This

makes SVD an easy target even for weak adversaries (e.g.,

non-adaptive), against which our algorithms are immune.

Best value approaches. Much of the research on rec-

ommendation systems concentrated on a passive model,

where the algorithm analyzes the data and is supposed

to make recommendations; the influence of the algorithm

on what data is collected, was ignored. An active model

of recommendation systems was introduced in [5] (called

“competitive recommendation systems” there). In this

model the objective is to find, for each user, at least one

object he rates as “good.” If preference vectors of differ-

ent types are orthogonal (i.e., do not share objects they

both rate as “good”), then these algorithms can be used to

reconstruct the complete preference vectors of the users.

In fact, [5] uses SVD and works only if type orthogonality

holds; it is vulnerable to attacks by dishonest users. In [2],

it is shown that a simple committee-based algorithm (see

below) suffices to find a good recommended object with-

out restricting the user preferences. Additionally, [2] give

a distributed peer-to-peer algorithm for picking a good

object, that can withstand any number of dishonest users

(the performance depends only logarithmically on the to-

tal number of users).

Again, our algorithms find the complete preference

vectors without assumptions. We also note that finding

complete vectors is useful in the long run: as a by-product,

our algorithms generate a classification of users into types,

which can be re-used when coping with a new batch of

objects.

Learning relations.

learning model of Goldman et al. [6], where the algorithm

works for all inputs. The setting in [6] is that a central-

ized algorithm needs to learn a binary relation: the rela-

tion value for all (i,j) pairs must be output. In each step,

the algorithm predicts the value of the relation for some

pair i,j, and then the true value is revealed. The measure

of performance in this case is the number of errors com-

mitted by the algorithm. For this model they consider a

few types of schedules, including the case where the algo-

rithm chooses which entry to guess next, which is similar

to the on-line model we consider. However, there are a

few important differences between the learning model and

ours. First, since the true values become known while

the algorithm is running, the task is easier than the one

we consider, where almost all values remain unexposed

even after the algorithm has terminated. Formally, in our

model, revealing any true grade incurs cost, and hence

the cost of the algorithm of [6] in our model is linear, i.e.,

the worst possible. Second, the learning algorithms are

fundamentally centralized in the sense that some play-

ers (possibly all) will probe all objects similarly to the

committee-based algorithms (see below).

Another related problem is the

5.1Asymmetric algorithms

One simplistic approach to solve the recommendation

problem is to use committees [5, 2]. In this case, a few

players are essentially charged with doing the work for all

others. Consider the following algorithm for our problem:

1. Choose K = O

?log n

α

?random players.

2. Let each player probe all objects.

3. Let all other players find the committee member

with whom they agree.

Step 3 can be implemented in K probes by each non-

committee player (as done in Steps 5–6 of Algorithm D in

Section 4). The correctness of the algorithm follows from

the observation that with high probability, the commit-

tee contains at least one member from each type whose

frequency is at least α. However, as mentioned in the

Page 7

introduction, committee-based algorithms are simply un-

acceptable in a distributed setting, due to their inherent

unfairness, and to their vulnerability to malicious attacks.

6.Conclusion and Open Prob-

lems

In this paper we showed that in a highly simplified

model, there exist very simple solutions to the impor-

tant problem of low-cost preference reconstruction. Our

results allow us to hope that the considerable gaps be-

tween the abstract model and real life can be bridged by

algorithmic solutions. In particular, we believe that the

following problems are important to solve.

• The algorithms in this paper are synchronous, i.e.,

players are assumed to proceed in lockstep fashion.

Finding an asynchronous algorithm would be a sig-

nificant progress in making the algorithm practical.

• Another important aspect is that the algorithms

in this paper assume that all players of the same

type have identical preference vector. It would be

extremely useful to extend the algorithms to the

case where types are not sharply defined, i.e., to

allow some tolerance for deviation.

• The algorithms use random player and product

identities. It seems interesting to try to develop

an algorithm that lifts this assumption.

• The algorithms rely on knowing a lower bound on

α, the popularity of the type. We would like to find

an algorithm that adapts to any α.

Taste and honesty. One (post modern) interpretation for

taste is honesty. In this model, each object has a single

value (“truth”), independent of the player who probes it.

This model is applicable if the grades of objects reflect

some objectively defined predicate. In this case a player

who announces a value which is not the true value of the

object is simply lying. With this interpretation, our algo-

rithms can be used to ensure that, with high probability,

each and every lie will eventually be uncovered. This cer-

tainty may lead to the following intriguing consequence

(assuming that the penalty for getting caught lying is

larger than the gain): no rational player will ever lie, and

the overhead of enforcing honesty will be minimized, since

the algorithm will be invoked only in the rare cases of ir-

rationality. Analyzing this scenario is beyond the scope

of the current paper.

Acknowledgment

We thank Avrim Blum for pointing the existence of [6] to

us.

References

[1] B. Awerbuch, B. Patt-Shamir, D. Peleg, and

M. Tuttle. Collaboration of untrusting peers with

changing interests. In Proc. 5th ACM Conf. on

Electronic Commerce (EC), pages 112–119, May

2004.

[2] B. Awerbuch, B. Patt-Shamir, D. Peleg, and

M. Tuttle. Improved recommendation systems. In

Proc. 16th ACM-SIAM Symposium on Discrete

Algorithms (SODA), pages 1174–1183, January

2005.

[3] Y. Azar, A. Fiat, A. Karlin, F. McSherry, and

J. Saia. Spectral analysis of data. In Proc. 33rd

ACM Symp. on Theory of Computing (STOC),

pages 619–626, 2001.

[4] J. F. Canny. Collaborative filtering with privacy. In

IEEE Symposium on Security and Privacy, pages

45–57, 2002.

[5] P. Drineas, I. Kerenidis, and P. Raghavan.

Competitive recommendation systems. In Proc.

34th ACM Symp. on Theory oand that the identities

of objects form a random permutation off

Computing (STOC), pages 82–90, 2002.

[6] S. A. Goldman, R. L. Rivest, and R. E. Schapire.

Learning binary relations and total orders. SIAM J.

Computing, 22(5):1006–1034, October 1993.

[7] J. Kleinberg and M. Sandler. Convergent algorithms

for collaborative filtering. In Proc. 4th ACM Conf.

on Electronic Commerce (EC), pages 1–10, 2003.

[8] R. Kumar, P. Raghavan, S. Rajagopalan, and

A. Tomkins. Recommendation systems: A

probabilistic analysis. In Proc. IEEE Symp. on

Foundations of Computer Science (FOCS), pages

664–673, 1998.

[9] S. K. Lam and J. Riedl. Shilling recommender

systems for fun and profit. In 13th conference on

World Wide Web, pages 393–402. ACM Press, 2004.

[10] R. Motwani and P. Raghavan. Randomized

Algorithms. Cambridge University Press, 1995.

[11] M. P. OMahony, N. J. Hurley, and G. C. M.

Silvestre. Utility-based neighbourhood formation for

efficient and robust collaborative filtering. In 5th

ACM conference on Electronic Commerce, pages

260–261, 2004.

[12] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom,

and J. Riedl. Grouplens: an open architecture for

collaborative filtering of netnews. In 1994 ACM

conference on Computer supported cooperative work,

pages 175–186. ACM Press, 1994.

[13] B. Sarwar, G. Karypis, J. Konstan, and J. Reidl.

Item-based collaborative filtering recommendation

algorithms. In 10th international conference on

World Wide Web, pages 285–295. ACM Press, 2001.

[14] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl.

Analysis of recommendation algorithms for

e-commerce. In 2nd ACM conference on Electronic

Commerce, pages 158–167. ACM Press, 2000.

[15] A. I. Schein, A. Popescul, L. H. Ungar, and D. M.

Pennock. Methods and metrics for cold-start

recommendations. In 25th annual international

ACM SIGIR conference on Research and

development in information retrieval (SIGIR ’02),

pages 253–260, 2002.

#### View other sources

#### Hide other sources

- Available from Zvi Lotker · May 31, 2014
- Available from Zvi Lotker · May 31, 2014
- Available from psu.edu