Triplet Losses-based Matrix Factorization for Robust Recommendations

2025-05-06 0 0 656.2KB 6 页 10玖币

侵权投诉

Triplet Losses-based Matrix Factorization for Robust

Recommendations

Flavio Giobergia1

1Department of Control and Computer Engineering, Politecnico di Torino, Turin, Italy

Abstract

Much like other learning-based models, recommender systems can be aected by biases in the training data. While typical

evaluation metrics (e.g. hit rate) are not concerned with them, some categories of nal users are heavily aected by these

biases. In this work, we propose using multiple triplet losses terms to extract meaningful and robust representations of users

and items. We empirically evaluate the soundness of such representations through several “bias-aware” evaluation metrics, as

well as in terms of stability to changes in the training set and agreement of the predictions variance w.r.t. that of each user.

Keywords

recommender systems, matrix factorization, contrastive learning

1. Introduction

Recommender systems are a fundamental part of almost

any experience of online users. The possibility of rec-

ommending options tailored to each individual user is

one of the key contributors to the success of many com-

panies and services. The metrics that are commonly

used in literature to evaluate these models (e.g. hit rate)

are typically only concerned with the overall quality of

the model, regardless of the behaviors of such models

on particular partitions of data. This results in recom-

mender systems typically learning the preferences of the

“majority”. This in turn implies a poorer quality of recom-

mendations for users/items that belong to the long tail

of the distribution. In an eort to steer the research fo-

cus to addressing this problem, the EvalRS challenge [

This challenge, based on the RecList framework [

], pro-

poses a recommendation problem with a multi-faceted

evaluation, where the quality of any solution is not only

evaluated in terms of overall performance, but also based

on the results obtained on various partitions of users and

items. In this paper, we present a possible recommender

system that addresses the problem proposed by EvalRS.

The solution is based on matrix factorization by fram-

ing an objective function that aligns users and items in

the same embedding space. The matrices are learned by

minimizing a loss function that includes multiple triplet

losses terms. Dierently from what is typically done (i.e.

aligning an anchor user to a positive and a negative item),

in this work we propose additionally using triplet terms

for users and items separately.

The full extent of the challenge is described in detail in

[

]. In short, the goal of the challenge is to recommend

EvalRS at CIKM 2022

"avio.giobergia@polito.it (F. Giobergia)

0000-0001-8806-7979 (F. Giobergia)

Attribution 4.0 International (CC BY 4.0).

CEUR

Workshop

Proceedings

http://ceur-ws.org

ISSN 1613-0073

CEUR Workshop Proceedings (CEUR-WS.org)

songs to a set of users, given their previous interactions

with other songs. The provided dataset is based on a

subset of the openly available LFM-1b dataset [

]. The

source code for the proposed solution has been made

available on GitHub 1.

2. Methodology

In this section we present the proposed methodology,

highlighting the main aspects of interest. No data prepro-

cessing has been applied to the original data, although

some approaches have been attempted (see Section 4).

The proposed methodology, as explained below, allows

ranking all items based on estimated compatibility with

any given user. We produce the nal list of

𝑘

recommen-

dations by stochastically selecting items from the ordered

list of songs, weighting each song with the inverse of its

position in the list.

2.1. Loss definition

Matrix factorization techniques have long been known

to achieve high performance in various recommendation

challenges [

]. This approach consists in aligning vec-

tor representations for two separate entities, users and

items (songs, in this case). This alignment task is a recur-

ring one: a commonly adopted approach to solving this

problem is through the optimization of a triplet loss [5].

A triplet loss is a loss that requires identifying an an-

chor point, as well as a positive and a negative point, i.e.

points that should either lie close to (positive) or far from

(negative) the anchor point.

Users and songs can thus be projected to a common

embedding space in a way that users are placed close to

songs they like and away from songs they do not like.

1https://github.com/fgiobergia/CIKM-evalRS-2022

arXiv:2210.12098v1 [cs.IR] 21 Oct 2022

This can be done by choosing a user as the anchor, and

two songs as the positive and negative points. A reason-

able choice for the positive song is one that has been

listened by the user. The choice for the negative song

is trickier. Random songs, or songs not listened by the

user are possible choices. However, more sophisticated

strategies can be adopted to choose negative points that

are dicult for the model to separate from the anchor.

These are called hard negatives and have been shown in

literature to be benecial to the training of models [6].

We decided to use a simple policy for the selection of

a negative song: a negative song for user

𝑢

is extracted

from the pool of songs that have been listened by one of

the nearest neighbors of

𝑢

and have not been listened by

𝑢

. By doing so, we aim to reduce the extent to which the

model relies on other users’ preferences to make a recom-

mendation. The concept of neighboring users is obtained

by comparing the similarity between embedding repre-

sentations of all users. Due to the computational cost of

this operation, it is only performed at the beginning of

each training epoch.

We can thus dene the triplets

(𝑢𝑎

𝑖, 𝑠𝑝

𝑖, 𝑠𝑛

𝑖)

to be used

for the denition of a triplet loss. Here,

𝑢𝑎

𝑖

is used to

represent the vector for the anchor user, whereas

𝑠𝑝

𝑖

and

𝑠𝑛

𝑖

represent the vectors for the positive and negative

songs respectively.

Similar approaches where users are aligned to songs

they did or did not like are Bayesian Personalized Rank-

ing (BPR) [

], where negative songs are sampled ran-

domly, and WARP [

], where negative items are sampled

so as to be “hard” (based on their proximity of the anchor

w.r.t. the positive item). To improve the robustness of the

representations built, we are additionally interested in

aligning similar songs and similar users. To this end, we

introduce two additional triplet terms to the loss function,

one that is based on

(𝑠𝑎

𝑖, 𝑠𝑝

𝑖, 𝑠𝑛

𝑖)

and one on

(𝑢𝑎

𝑖, 𝑢𝑝

𝑖, 𝑢𝑛

𝑖)

Based on the previously dened concepts, we choose

𝑠𝑎

𝑖

as a song listened by

𝑢𝑎

𝑖

, and

𝑢𝑝

𝑖

and

𝑢𝑛

𝑖

as users who

respectively listened to

𝑠𝑝

𝑖

and

𝑠𝑛

𝑖

. Other alternatives

have been considered, but were ultimately not selected

due to a higher computational cost.

We dene the nal loss as:

ℒ=

𝑖

𝑤𝑖(𝑚𝑎𝑥{𝑑(𝑢𝑎

𝑖, 𝑠𝑝

𝑖)−𝑑(𝑢𝑎

𝑖, 𝑠𝑛

𝑖) + 𝑚0,0}+

𝜆1𝑚𝑎𝑥{𝑑(𝑢𝑎

𝑖, 𝑢𝑝

𝑖)−𝑑(𝑢𝑎

𝑖, 𝑢𝑛

𝑖) + 𝑚1,0}+

𝜆2𝑚𝑎𝑥{𝑑(𝑠𝑎

𝑖, 𝑠𝑝

𝑖)−𝑑(𝑠𝑎

𝑖, 𝑠𝑛

𝑖) + 𝑚2,0})

(1)

Where

𝑑(·)

is a distance function between any pair of

vectors. In this work, the cosine distance is used.

𝑚𝑗

is a

margin enforced between positive and negative pairs. In

this work, since all elements are projected to a common

embedding, we used

𝑚0=𝑚1=𝑚2

. Finally,

𝑤𝑖

is a

weight that is assigned to each entry, which is discussed

useranc

songanc

userneg

songneg

userpos

songpos

user-song loss

song-song loss

user-user loss

Figure 1:

Representation of the action of each part of the loss

on the vectors. Arrow directions represent whether elements

are pulled towards or pushed away from the anchors.

in Subsection 2.2.

Figure 1summarizes the eect of the various terms of

the loss on the embedding vectors learned.

2.2. Popularity weight

To make the minority entities more relevant, we adopted

a weighting scheme that modulates the previously de-

scribed loss so as to weigh rows more if they belong to

“rarer” entities and less for common ones. In accordance

with [

], we identied ve factors to be kept into account.

Based on these, a coecient has been dened for each

entry in the training set. The nal weight is given by

a weighted average of these coecients. The following

is a list of factors, along with the way the respective co-

ecients have been computed (logarithms are used for

factors that follow a power law distribution). All coe-

cients are normalized to sum to 1 across the respective

population.

•

Gender (

𝜃𝑔𝑒𝑛𝑑𝑒𝑟

): in accordance with the original

dataset, a relevance coecient is provided for the

categories male, female, and undisclosed

. The

coecient is proportional to the inverse of the

occurrences of each gender in the list of known

users.

This simplied perspective on gender does not reect that of the

author

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TripletLosses-basedMatrixFactorizationforRobustRecommendationsFlavioGiobergia11DepartmentofControlandComputerEngineering,PolitecnicodiTorino,Turin,ItalyAbstractMuchlikeotherlearning-basedmodels,recommendersystemscanbeaffectedbybiasesinthetrainingdata.Whiletypicalevaluationmetrics(e.g.hitrate)arenotc...

展开>> 收起<<

Triplet Losses-based Matrix Factorization for Robust Recommendations.pdf

共6页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Triplet Losses-based Matrix Factorization for Robust Recommendations

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: