Automatic Generation of Product Concepts from Positive Examples with an Application to Music Streaming

2025-05-02 0 0 1.61MB 18 页 10玖币

侵权投诉

Automatic Generation of Product Concepts

from Positive Examples, with an Application to

Music Streaming

Kshitij Goyal1, Wannes Meert1, Hendrik Blockeel1, Elia Van Wolputte1, Koen

Vanderstraeten2, Wouter Pijpops2, and Kurt Jaspers2

1KU Leuven, Belgium

2Tunify, Belgium

Abstract. Internet based businesses and products (e.g. e-commerce,

music streaming) are becoming more and more sophisticated every day

with a lot of focus on improving customer satisfaction. A core way they

achieve this is by providing customers with an easy access to their prod-

ucts by structuring them in catalogues using navigation bars and provid-

ing recommendations. We refer to these catalogues as product concepts,

e.g. product categories on e-commerce websites, public playlists on music

streaming platforms. These product concepts typically contain products

that are linked with each other through some common features (e.g. a

playlist of songs by the same artist). How they are deﬁned in the back-

end of the system can be diﬀerent for diﬀerent products. In this work,

we represent product concepts using database queries and tackle two

learning problems. First, given sets of products that all belong to the

same unknown product concept, we learn a database query that is a

representation of this product concept. Second, we learn product con-

cepts and their corresponding queries when the given sets of products

are associated with multiple product concepts. To achieve these goals,

we propose two approaches that combine the concepts of PU learning

with Decision Trees and Clustering. Our experiments demonstrate, via

a simulated setup for a music streaming service, that our approach is

eﬀective in solving these problems.

Keywords: PU Learning ·Machine Learning ·Music Streaming

1 Introduction

Machine learning is used for various applications these days and more and more

businesses are looking to use machine learning to improve their products. Re-

cent advances in online consumer based businesses provide an opportunity to

explore machine learning solutions to challenging problems. One such problem

is generating dynamic product concepts that contain a selection of products that

are linked with each other through some common features. For example, in e-

commerce a product category that contains similar items (e.g., ‘cosmetic items’)

is a product concept, in music streaming services a public playlist is a product

arXiv:2210.01515v3 [cs.LG] 15 Jan 2023

2 K. Goyal et al.

concept. Typically these kinds of product concepts are available for users to se-

lect from. There are beneﬁts of having such product concepts in your system: 1.

they provide a good way to structure the itinerary; 2. a user can use them to

navigate the website; 3. they make it easy for users to discover new items.

Due to the lack of transparency behind how these product concepts are cre-

ated for most services, we can not generalize their creation in the back-end of

a system. However, in our work, we assume that a product concept is associ-

ated with a database query, which we term as the concept query, that ﬁlters

that whole database of items (products, songs etc.) based on certain common

features. This makes these product concepts dynamic in nature: they get auto-

matically updated when new items are added to the database. This deﬁnition is

inspired by our use case of a music streaming company called Tunify3.

Tunify is a music streaming service that provides a predeﬁned selection of

playlists to businesses. Tunify has a database of songs where each song is rep-

resented by a ﬁxed set of discrete valued features: mood, popularity etc. Tunify

also maintains a set of database queries that deﬁne useful product concepts

(the product concept deﬁned by a query is a set of all songs that are returned

when that query is run on the database). These products concepts are useful

for generating playlists for businesses. A business can select a product concept

based on a small description (e.g., 80’s Rock) and a playlist based on the se-

lected product concept is generated, the generated playlist is a sample of all the

songs associated with the product concept. The database queries corresponding

to product concepts are manually deﬁned by music experts that Tunify employs.

This query creation process has an obvious drawback: it requires a lot of time to

ﬁne tune the exact feature-values the query should contain. This motivates our

ﬁrst problem: can we automatically identify the database query corresponding

to a product concept if we are provided with a set of playlists that the target

product concept should generate? We argue that it is easier for an expert to

manually create playlists that the target concept should generate compared to

manually creating a database query.

Another interesting problem setting is when the provided playlists come from

multiple target product concepts. Can we identify diﬀerent concepts the playlists

are coming from and their corresponding database queries? This problem is mo-

tivated by the fact that Tunify allows for customers to create their own playlists

under one of their subscriptions. As these playlists come from multiple customers,

we expect that: 1. not all of them contain similar songs; 2. there are multiple

playlists that contain similar songs. We want to identify similar playlists and

create product concepts based on them. There are two outcomes of this: either

the identiﬁed product concepts are missing from Tunify’s system, or the iden-

tiﬁed product concepts are already in the system but the customers didn’t use

them for whatever reasons. In the former case, we improve the database with

new concepts, and in the later case, Tunify could reach out to the customers

that created their own playlists and recommend the already existing concepts.

3https://www.tunify.com/nl-be/language/

Automatic Generation of Product Concepts from Positive Examples 3

Even though we motivate these problems based on our use case of Tunify,

these problems are generally applicable for any business where such product

concepts are used. Consider the example of an e-commerce company, here the

product concepts are the product categories (e.g., cosmetics, menswear). The

customers interact with the product concept in this case diﬀerently from Tunify:

in Tunify, a playlist is generated when a customer selects a product concept,

but in the case of e-commerce a customer can view all the products associated

with a product concept. In the context of our two learning problems, however,

this diﬀerence does not have any impact. For the ﬁrst learning problem, instead

of experts making a playlist of songs, the experts create a set of items. For the

second problem, instead of using playlists of songs created by customers, we can

use the items the customers purchased together.

To summarize, we consider two learning problems: 1. learning the product

concept query given a collection of itemsets that are associated with it, 2. learn-

ing product concepts and their corresponding queries given a collection of item-

sets that may or may not be associated with a single product concept. For the

ﬁrst problem, we use a combination of PU learning (Positive and Unlabelled

Learning) [4] techniques with decision tree learning to learn the product concept

queries as the rules from the decision tree. In addition to this, we also study the

eﬀect of noise in the provided itemsets on the ﬁnal query and propose a way to

deal with it. For the second problem, we combine the approach of task one with

clustering to identify new product concepts from data generated by customers.

We design and test our experiments on the dataset provided to us by Tunify.

With a simulated experimental setup, we demonstrate that our approaches are

able to learn good quality concept queries with small number of items, even when

there is noise in the set, and we are able to eﬀectively identify product concepts

using the customer data. We additionally show that our proposed algorithm for

the ﬁrst problem is robust to noise in the provided set of items.

The paper is structured in the following way: ﬁrst we introduce some termi-

nologies and explain the problem statements in section 2, secondly we present

our approach in section 3, then we present the experimental results in section 4

before the related works and a discussion in sections 5 and 6 respectively. Section

7 concludes.

2 Framework

In this section, we ﬁrst give an overview of some concepts from logic and satis-

ﬁability which we use in our work before explaining our problem statement.

2.1 Propositional Logic

Propositional Logic formulas contain literals which are Boolean formulas, their

negation and logical connectives, e.g., pa_ pb^ pcqqq. An assignment xof vari-

ables ta, b, cusatisﬁes a formula φif xmakes the formula φTrue. Any logical

formula can be rewritten in a normal form such as Conjunctive Normal Form

4 K. Goyal et al.

(CNF) or Disjunctive Normal Form (DNF). A CNF formula consists of con-

junction of disjunction of literals and a DNF formula contains disjunction of

conjunction of literals, where conjunction is the logical ‘AND’ (^) operator and

disjunction is the logical ‘OR’ (_) operator.

2.2 Problem Statement

We now formally deﬁne our learning problems. The dataset of instances is repre-

sented by D. We assume that an instance is represented by a ﬁxed set of discrete

valued features Fand takes a single value for each feature. For a feature fPF,

the discrete set of values fcan take is represented by Vpfq.

Deﬁnition 1. Product Concept. A product concept is a collection of instances.

A product concept is associated with a concept query that deﬁnes which instances

belong to it.

A product concept can be a union of multiple ‘sub-concepts’. For example,

in e-commerce, a category can have many diﬀerent sub categories; in Tunify,

there are a number of product concepts that combine the music from multiple

diﬀerent product concepts to generate a playlist that contains songs from all

the combined product concepts (e.g., ‘Fitness Center’ product concept combines

product concepts ‘Dance Workout’ and ‘Rock Dynamic’). Keeping this in mind,

we formally deﬁne a concept query as follows:

Deﬁnition 2. Conjunctive Concept Query. Given a set of attributes FĎF

and sets of values VfĎVpfqfor each fPF. A conjunctive concept query Q,

for an arbitrary input xPD, is deﬁned as the following rule-based query in a

conjunctive normal form:

Q:ľ

fPF

vPVf

pxf“vq

Deﬁnition 3. Concept Query. A concept query is deﬁned as:

1. A conjunctive concept query is a concept query.

2. A disjunction of two or more conjunctive concept queries is a concept query.

In the case where a concept query is a disjunction of two or more conjunctive

concept queries, each conjunctive concept query and all the items that make it

true are said to be associated with a sub-concept of the parent product concept

(where the parent product concept is the disjunctive combination of the sub-

concepts). Any item that is associated with any of the sub-concepts is said to

belong to the parent product concept. By deﬁnition, a sub-concept is also a

product concept. We will refer to the concept query corresponding to a product

concept Cas QCin the text from now on. Also, for a given query Q, the items

from the database Dthat are ﬁltered by the query are denoted by QpDq.

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：

AutomaticGenerationofProductConceptsfromPositiveExamples,withanApplicationtoMusicStreamingKshitijGoyal1,WannesMeert1,HendrikBlockeel1,EliaVanWolputte1,KoenVanderstraeten2,WouterPijpops2,andKurtJaspers21KULeuven,Belgium2Tunify,BelgiumAbstract.Internetbasedbusinessesandproducts(e.g.e-commerce,musicstr...

展开>> 收起<<

Automatic Generation of Product Concepts from Positive Examples with an Application to Music Streaming.pdf

共18页,预览4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！

Automatic Generation of Product Concepts from Positive Examples with an Application to Music Streaming

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: