Knowledge Graph-based Product Recommendations on e-Commerce

Platforms

Andr

e Gomes Regino

1 a

, Rodrigo Oliveira Caus

, Victor Hochgreb

1 b

and Julio Cesar Dos Reis

2 c

Institute of Computing, University of Campinas, Campinas, S

ao Paulo, Brazil

GoBots, Campinas, S

ao Paulo, Brazil

Keywords:

Recommendation Systems, Knowledge Graphs, Question Answering Systems.

Abstract:

The amount of data generated in e-commerce sales has expressively grown in the last few years. Online stores

often receive questions about products related to price, guarantee, and shipping price. By reducing time for

prompt answering, stores can improve customer satisfaction and sales conversion rate. The recommendation

of available alternative products in case of product unavailability intended by the customer plays a key role in

sales growth in this context. This article deﬁnes and evaluates a technique for product recommendation based

on the product’s facts stored in Knowledge Graphs (KGs). Our KG is ﬁlled with facts from natural language

questions and answers processed from the e-commerce platform. We exemplify our proposal in a real-world

solution, using data from online stores processed by GoBots, a leading e-commerce chatbot business in Latin

America. Online sellers assessed the results of the recommendations to evaluate their quality.

1 INTRODUCTION

The convenience of online stores has captured cus-

tomers’ attention who, a few decades ago, only made

face-to-face purchases in commercial establishments.

The COVID-19 pandemic has changed customers’ re-

lationships with such online stores because they were

one of the few ways to buy products. The virtual

stores, called e-commerces, had to adapt to this grow-

ing demand. In this context, the guarantee of security

and integrity in online sales should be allied to fair

prices and deliveries in a reasonable time. These as-

pects guarantee a better customer experience through-

out the buying process.

Product recommendation ﬁgures as a key strategy

to offer a complete service to the ﬁnal customer. This

aims to deliver one or more products that suit the cus-

tomer’s taste or need. The suggested product should

be compatible with the customer’s purchasing behav-

ior or characteristics. An adequate recommendation

is an ally when the required product is not in stock

or is no longer traded. In this case, a sale compati-

ble with the customer’s needs would please both the

customer and the seller. In this scenario, the customer

https://orcid.org/0000-0001-9814-1482

https://orcid.org/0000-0002-0529-7312

https://orcid.org/0000-0002-9545-2098

would not leave without the purchase, and the seller

would not miss a sale.

The recommendation system functions as an in-

formation ﬁltering. In the e-commerce scenario, the

aim is to ﬁlter products that may be interesting to a

given customer (Shao et al., 2021), choosing prod-

ucts when there are many available options (Isinkaye

et al., 2015). Amazon, a big player in the e-commerce

scenario, increased its sales by 35% after adopting a

recommendation system to display speciﬁc products

to its customers (Lee and Hosanagar, 2014). How-

ever, recommending a product is not an easy task. It

depends on the recommendation strategy adopted to

be successful.

Existing approaches use different data sources

to identify the compatibility between the customer

purchase history and the catalog of items in the e-

commerce platforms. Examples of these data are

product ratings, attributes, and user search history

(Dwivedi et al., 2020). The strategy can vary from

collaborative ﬁltering (Linden et al., 2003), machine

learning (Covington et al., 2016) and Knowledge

graph-based techniques (Guo et al., 2020). Lately,

knowledge graphs (KGs) have been studied and ex-

plored for recommendation purposes. For example,

AliCoco, a KG used in the largest e-commerce in

China, AliBaba (Luo et al., 2020). KGs are helpful to

represent connections between customers, products,

Regino, A., Caus, R., Hochgreb, V. and Reis, J.

Knowledge Graph-based Product Recommendations on e-Commerce Platforms.

DOI: 10.5220/0011388300003335

In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 2: KEOD, pages 32-42

ISBN: 978-989-758-614-9; ISSN: 2184-3228

and sales.

This article proposes a novel methodology rely-

ing on KGs to recommend products based on the

compatibility between an item owned by the cus-

tomer and the product being sold. Our solution differs

from the state-of-art studies. Our approach performs

product recommendations searching for compatibil-

ity connections represented (encoded) in a KG. The

recommendation is based on a compatibility question

made by the customer and the compatibility among

the desired product (by the customer) and similar

products indexed by our KG triples. No user data

is required in our proposed recommendation process.

Our KG was initially intended to answer compatibil-

ity questions made by customers in large Brazilian e-

commerce platforms (Sant’Anna et al., 2020).

The contribution of this research is to further ex-

tend the system functionality by automatically pro-

viding answers with recommended products in case

of incompatibility between the item owned by the

customer and the desired product. To the best of

our knowledge, there is no evidence in literature

of a solution that recommends products in an e-

commerce context based on a question-answering ap-

proach combined with the products’ attributes rep-

resentation. Our evaluation results based on sellers’

opinions reveals that the suggested recommendations

from the system are adequate. We experienced an in-

crease in the number of recommendations provided

by the system.

The remaining of this article is organized as fol-

lows: Section 2 discusses related work; Section 3

presents our proposed solution by including the for-

malization of our recommendation technique and a

running example from a real-world e-commerce plat-

form. Section 4 conducts an evaluation to assess sell-

ers’ perspectives and perceptions on the recommen-

dation features and results in the platform. In section

5 we further discuss our ﬁndings. Section 6 presents

the ﬁnal remarks.

2 RELATED WORK

This section presents relevant approaches from state-

of-the-art studies of knowledge-graph-based recom-

mendation systems.

Guo et al. (Guo et al., 2020) gathered a set

of 39 studies focused on KG-based recommendation

systems. They classiﬁed the recommendation algo-

rithms into three groups: the collaborative-ﬁltering-

based, the content-based, and the hybrid solutions.

The collaborative-ﬁltering aims to match the user pro-

ﬁle with other proﬁles in the platform (e-commerce,

video streaming, etc), matching their behavior. It is

the most common recommendation method. In or-

der to improve its efﬁciency and accuracy, several

studies explored external sources to gather user data,

such as social media and user location. The content-

based approach uses past users’ information and be-

havior tracking in the platform. The rationale is that

users may be interested in products similar to those

they already bought. The hybrid category addresses

challenges found in both collaborative-ﬁltering and

content-based categories: the cold start, data spar-

sity, and heterogeneity, to cite a few. The authors

also discussed a list of the main challenges in the

relationship between KGs and recommendation sys-

tems. The time-consuming is a relevant aspect to take

into account. These methods use machine learning

algorithms, such as Convolutional Neural Networks

(CNN), to train the recommendation model. In large

online systems, the recommendations should be up-

dated constantly, requiring a fully retraining of the

deployed models. The authors indicated this task of

model training and its overtime update as one of the

signiﬁcant challenges in the area.

AliCoCo (Luo et al., 2020) is a state-of-art KG

that conceptualizes users’ needs in Alibaba’s plat-

form. Their motivation relied on users’ dissatisfaction

with Alibaba’s recommendation results, which gener-

ated much redundant display of products. To over-

come it, a KG was built to encode brands, categories

of products, and search concepts. Their study (Luo

et al., 2020) found that customers in Alibaba’s plat-

form are interested in concepts like “Outdoor Fam-

ily Barbecue”. This consists of products of different

types and brands (grill, knife, butter, sunblock), not

only individual products. The reported results indi-

cated an increase in the click-through rate and the cus-

tomer satisfaction based on GMV (Gross Merchan-

dise Value) rate.

Wu et al. (Wu et al., 2019) presented a summary

of methods related to how natural language ques-

tions are answered based on KGs. The authors or-

ganized existing KGs by their domain application like

open domain KGs, functioning as an encyclopedia for

several domains (e.g., DBpedia) and domain-speciﬁc

KGs. We classify our solution as a niched category at

the current stage, modeling the automobile domain.

This speciﬁc domain was considered in our study due

to the high amount of existing compatibility questions

in the e-commerce platform. The RDF triples in our

KG were constructed based on the processing of these

questions (Sant’Anna et al., 2020).

In addition, according to Wu et al. (Wu et al.,

2019), the second category organization relates to

KGs parsing method by classifying them in two

Knowledge Graph-based Product Recommendations on e-Commerce Platforms

groups: semantic parsing and information retrieval.

In semantic parsing, the questions are transformed

into logical symbols to represent the semantics in-

tended by the authors’ questions. The disadvantage

of semantic parsing is that it is highly dependent on

external resources, such as dictionaries and labeling

techniques, to categorize data modeled by the KG

(Wu et al., 2019). On the other hand, the informa-

tion retrieval extracts relevant parts of the questions

to query the KG looking for candidate answers. After

ranking the found answers, it creates an optimized ﬁ-

nal answer. Wu et al. (Wu et al., 2019) described that

this technique has several performance issues when

compared to semantic parsing methods.

Our proposed recommendation system is suited

to automatically recommend products (represented in

our RDF KG) in an e-commerce platform. We im-

plemented it in a real-world software application run-

ning for online stores. Our main contribution refers

to an original recommendation technique fully de-

veloped that explores compatibility facts encoded in

our KG to produce recommendations based on prod-

ucts’ attributes. We deployed and applied our sys-

tem to an e-commerce platform in Brazil. To the best

of our knowledge, there is no evidence in the liter-

ature of a solution that recommends products in an

e-commerce context based on a question-answering

approach combined with the products’ attributes rep-

resentation.

3 RECOMMENDATION SYSTEM

FOR E-COMMERCE

This section describes an overview of our recommen-

dation system (cf. Subsection 3.1). Subsection 3.2

further formalizes and presents details of our speciﬁc

KG-based recommendation used to suggest products

in e-commerce platforms. Subsection 3.3 shows a

running example of a recommendation proposed by

our system.

3.1 Overview

A KG is a directed graph with nodes representing

real-world entities such as “Rio de Janeiro”; and

edges representing relations between entities; e.g.,

“Rio de Janeiro” “countryName” “Brazil”. KGs have

the form of RDF triples in a way that G = t

,...,t

A RDF triple (t) refers to a data entity composed of

subject (s), predicate (p) and object (o) deﬁned as

t = (s, p,o).

Sant’Anna et al.(Sant’Anna et al., 2020) intro-

duced the KG used in this investigation. Our previous

work addressed how to generate triples by process-

ing pairs of customer’s questions and attendant’s an-

swers in natural language. The encoded knowledge

was used to automatically answer customer’s com-

patibility questions, indicating whether the product

ﬁts in their vehicles. This KG answered more than

31,000 compatibility questions of 77 stores in one of

the biggest Latin American marketplaces.

Our KG expresses knowledge on compatibility

between products of the automotive domain of e-

commerce and cars. Figure 1 shows an example of

a set of triples in our KG. It presents an instance of

the class Store (car-parts) that sells an instance of the

class Product (ID “108093”). A consumer item of the

class “Car” (“ﬁesta-2012”) is fully compatible with

the instance of the Product class.

Figure 2 presents the key steps in the recommen-

dation process designed in our approach based on a

compatibility question performed by a customer (for-

malized in Algorithm 1). The input is a compatibility

natural language question that a customer asks on a

product page. The output is an answer containing a

list of recommended products.

First, the e-commerce user (customer) asks a

question q written in a natural language text about a

product p. The question q is processed by a GoBots

service that implements an intent and entity discovery

process based on question q as input (step A in Figure

2).

At this step (Line 1 in Algorithm 1), the entities

E = {e

,..., e

} from the question are identiﬁed

(e.g., “brand” and “model” are used to identify at-

tributes of a car). Algorithm 1 identiﬁes the intention

i of the question (Line 2). Our recommendation al-

gorithm is suitable for questions of compatibility in-

tention (Line 3 in Algorithm 1). A customer asking

whether the product in question, which may be a car

part, ﬁts in a particular vehicle is a question with com-

patibility intent.

If the intention i is related to “compatibility” (Line

3 of Algorithm 1), our algorithm proceeds to the sec-

ond step (Step B in Figure 2). It evaluates whether

the item in the consumer’s possession ci (called “con-

sumer item” from now on) is compatible with the

product sold p (Line 5 of Algorithm 1).

The function “getCompatibility” queries the

KG retrieving a response r

. The response

= { f ,t,a

com

} from the KG is composed by a

boolean f , indicating whether the compatibility was

found or not; a compatibility type t and an attendant

answer a

com

. There are four types of valid compati-

bility, all of them representing disjoint subclasses of

the Compatibility KG class:

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

Figure 1: Example of a set of triples registered in our KG for complete compatibility facts between a product and the car

“Fiesta 2012”. The darker nodes represent classes and the lighter nodes represent instances.

T = {FullCompatibility,UniversalCompatibility,

ConditionalCompatibility,NoCompatibility}

(1)

The types of compatibility indicate that the prod-

uct ﬁts with the consumer item (Full Compatibility);

partially ﬁts (Conditional Compatibility); does not ﬁt

(No Compatibility); ﬁts with any consumer item (Uni-

versal Compatibility). An example of a Full Compat-

ibility type can be a given car exchange perfectly ﬁt-

ting in a “Honda Civic 2012”.

Algorithm 1: Product recommendation based on a con-

sumer item and a desired product.

Require: q,recommendedProducts, p

1: E ← identi f yEntities(q)

2: i ← identi f yIntent(q)

3: if i ≡ “compatibility

′′

then

4: ci ← extractConsumerItem(E)

5: r

← getCompatibility(ci, p)

6: if r

.t ≡ NoCompatibility then

7: recommendedProducts ←

Algorithm2(ci, p)

8: return recommendedProducts

9: else

10: return r

com

11: end if

12: end if

If t = NoCompatibility (Line 6), indicating that ci

is not compatible with p, Algorithm 1 proceeds to the

third step (output False of Step B in Figure 2 lead-

ing to Step C). At this point, a new request is per-

formed to the KG to identify which products from the

seller store P

rec

= {p

rec1

, p

rec2

,..., p

recn

} ﬁt the con-

sumer item ci (e.g., an older version of the car ex-

change may ﬁt in the Honda Civic 2012). Section 3.2

presents this procedure in more details. In case that

there are no available compatible products to be rec-

ommended, the procedure returns an answer stating

that ci ad p are not compatible (output False of step C

in Figure 2).

In case that t ̸= NoCompatibility (t =

FullCompatibility or t = ConditionalCompatibility

or t = UniversalCompatibility), the algorithm iden-

tiﬁes that the ci and p are compatible and returns

the answer that states the compatibility between the

items (Lines 9 and 10 of Algorithm 1).

3.2 Recommendation Technique

This subsection provides a deeper description of Step

C as the core of our recommendation system. In our

approach, the rationale is that ci is not compatible

with p, but there might be a set of products P

rec

pre-

sented in the KG that ﬁts ci. The difference between

the KG requests from steps B and C is that in B, our

algorithm veriﬁes the compatibility between the pair

(p, ci), whereas in step C, the request queries which

products p

rec

∈ P

rec

ﬁts ci. In addition, the customers

can be interested in other products of the same seller,

but they do not know which seller’s products ﬁt in

their consumer item.

The set of recommended products P

rec

is com-

posed based on inclusion criteria I. The motivation

is the necessity to retrieve a well-deﬁned set of avail-

able products, meeting a pre-deﬁned standard. The

following items show the criteria adopted by our al-

gorithm to recommend products:

Knowledge Graph-based Product Recommendations on e-Commerce Platforms

1. ∀ p

rec

∈ P

rec

and p should have the same category;

2. ∀ p

rec

∈ P

rec

should be compatible with consumer

item ci;

3. ∀ p

rec

∈ P

rec

and p should be offered in the same

store;

4. ∀ p

rec

∈ P

rec

should be available by any customer

to buy;

5. There must be one or more p

rec

∈ P

rec

;

The ﬁrst and second criteria show that the recom-

mended products should have the same category as

the product asked by the customer. They should ﬁt

in the consumer item described in the user’s question

(e.g., only car exchanges should be recommended to

the user if the compatibility question relates to a car

exchange ﬁtting a “Honda Civic 2012” car). The third

criterion is that only products sold by the seller should

be recommended. In an e-commerce platform, many

sellers are competing for a market share, and recom-

mending the competitor’s products would be unac-

ceptable. The fourth and ﬁfth criteria are the status of

the product campaign. In other words, if the product

is being sold and more than one unit of it is available.

Recommending a product out of stock could nega-

tively affect the seller’s reputation. The data required

in some items of I (category, store, product availabil-

ity) is retrieved using the seller’s marketplace APIs.

At Step C, Algorithm 2 requires ci, p and I, and

outputs an object of the class Recommendation (r),

composed by a boolean value indicating that the rec-

ommended products were found (r. f ound) and a rec-

ommendation answer (r.answer). Algorithm 2 uses a

SPARQL query (Line 2) to get a preliminary list of

products P

rec

registered in KG that have a FullCom-

patibility relation with ci. The Query 1 presents a

SPARQL template for the automotive domain. The

literal “PRODUCT

CATEGORY” represents the ti-

tle of the category attribute of p; the STORE NAME

resource represents the name of the store that sells

p; “CAR MODEL” and “CAR YEAR” literals repre-

sent the attributes model and year of ci, respectively.

Each resource or literal on this template is replaced

by its respective value obtained from p and ci.

BASE < ht t p :// kg . test / KB / k no w le d g e gr a ph

PREFIX on to : < htt p : / / kg . te st /KB /

ontol o gy >

SELECT ? id ? co m p a ti b i l it y FROM < h ttp ://

kg . tes t / KB >

WHERE {

? pr od uc t rdf : type o nto : P r od uc t .

? pr od uc t on t o : h a s Ca te go r y "

PR O D U C T _ CA T EG O RY " .

? pr od uc t on t o : has I D ? id .

? pr od uc t on t o : h as C o m p a t ib i li t y ?

co mp a ti bi l it y .

< ST O RE _ NAM E > o n to : se l ls ? pr o du ct .

{

? com p a t ib i l i ty onto :

co mp a ti b l e Wi t h ? ca r .

? car on to : h a sM od el " CAR _M OD EL "

? car on to : h as Mo d el Ye a r "

CA R _ Y E A R " .

? com p a t ib i l i ty rdf : type on t o :

Fu l l C o m p a t i b i l it y

}

Query 1: SPARQL template to retrieve P

rec

based on ci and

p category.

The list of recommended products retrieved by

Query 1 is ﬁltered based on the remaining items of

inclusion criteria I (Line 3 in Algorithm 2). After

querying and ﬁltering the set of products P

rec

, Algo-

rithm 2 applies a ranking strategy to delimit which

products might be more interesting for recommenda-

tion to the consumer. We explore the conversion rate

(conv). This rate is composed of the number of sales

(s) divided by the product’s visits (v). P

rec

is ordered

by conversion rate in descending order (i.e., the higher

the conversion, the higher the product is in the rank-

ing). The conversion rate is described in Equation 2.

At this stage, we could also rank products by sales,

but this would penalize recently added products in the

e-commerce platform.

conv(p) =

(

s(p)/v(p) if v > 0

0 otherwise

∀p ∈ P

rec

(2)

At this point, there is a set of ﬁltered and ranked

recommended products (Lines 3 and 6 in Algorithm

2). Additionally, we opted to limit the number of re-

sults. The top k products form the ﬁnal recommenda-

tion list stored in P

rec

. In case of recommended prod-

ucts (P

rec

̸=

0), our solution returns a range of recom-

mended products (Line 7 of Algorithm 2). The last

step is to build an answer r.answer containing the rec-

ommended products P

rec

links (Line 11 in Algorithm

2).

3.3 Running Example

This section presents a real-world running example

extracted from our KG (cf. Figure 3). Our KG is con-

stantly updated, reaching a top rate of 2,000 newly

added nodes per day. It has approximately 500,000

compatibility nodes and more than 3,500,000 triples.

This KG is used to answer real-time compatibility

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

Figure 2: Recommendation based on question-answer facts.

Algorithm 2: Recommendation procedure.

1: function recommendation(ci, p)

Require: ci, p, I, k

2: P

rec

← constructSparqlQuery(ci, p)

3: P

rec

← f ilterProducts(ci,I)

4: if P

rec

̸=

0 then

5: r. f ound ← True

6: P

rec

← rankProducts(P

rec

)

7: P

rec

← limitProducts(P

rec

,k)

8: else

9: r. f ound ← False

10: end if

11: r.answer ← buildAnswer(P

rec

)

12: return r

questions regarding automotive domain from cus-

tomers in a big Latin-American marketplace. In the

scenario described in Figure 3, a product p (speaker)

was found in an online automobile store in one of

the largest Latin America marketplaces. Initially, a

customer enters the speaker product page and asks a

question q. The customer is in doubt if the speaker

can be used in his car, a “Corsa 2004” model (ci). In

the Portuguese language, the customer creates a ques-

tion: “Can I use this speaker in my Corsa 2004 car?”

(Question translated to English by the authors).

This scenario shows an example of a question that

can result in a recommendation answer. The intent

(i) of the question is compatibility. “Corsa” is the car

model, and “2004” refers to the manufacturing year of

the car. Both characteristics form the deﬁnition of ci.

In this case, a car; and p is represented by a speaker in

stock. The question arrives at the GoBots service in a

textual format. The service is responsible for identi-

fying both i and E of q (Step A in Figure 2). The NLP

tool identiﬁes that the question has a compatibility in-

tent and some entities, including a model and a year

of a car product.

This data is sent to the KG service (Step B in Fig-

ure 2) responsible for handling events related to the

KG. A SPARQL query is formed to verify if there

is a compatibility between ci (“Corsa 2004”) and p

(“speaker”). This SPARQL query returns from the

KG whether the items are compatible. In our case,

the KG responded with compatibility type (t) equals

to “NoCompatibility”. Based on this response, our al-

gorithm proceeds to the recommendation step (C in

Figure 2). A new SPARQL query is formed, using as

parameter ci (“Corsa 2004”) and p (“speaker”).

Query 2 presents the recommendation query sub-

mitted to our KG related to our running example. It

selected products sold by a store (<car-sound-parts>)

presented in the same category of the speaker (“AU-

TOMOTIVE SPEAKER”) that are compatible with

a speciﬁc product (“car”), model (“Corsa”) and year

(“2004”).

BASE < ht t p :// kg . test / KB / k no w le d g e gr a ph

PREFIX on to : < htt p : / / kg . test / KB /

ontol o gy >

SELECT ? id ? co m p a ti b i l it y FROM < h ttp ://

kg . tes t / KB >

WHERE {

? pr od uc t rdf : type o nto : P r od uc t .

? pr od uc t on t o : h a s Ca te go r y "

Knowledge Graph-based Product Recommendations on e-Commerce Platforms

AU T OM O TI V E _ S P E A K E R " .

? pr od uc t on t o : has I D ? id .

? pr od uc t on t o : h as C o m p a t ib i li t y ?

co mp a ti bi l it y .

< st o re / car - sound - parts > ont o : s ell s

? pr od uc t .

{

? com p a t ib i l i ty onto :

co mp a ti b l e Wi t h ? ca r .

? car on to : h a sM od el " co rsa " .

? car on to : h as Mo d el Ye a r 200 4 .

? com p a t ib i l i ty rdf : type on t o :

Fu l l C o m p a t i b i l it y

}

Query 2: SPARQL query responsible for selecting

automotive speakers sold by the store compatible with

Corsa 2004

Figure 3: Running example of product recommendation.

The original question was written in Portuguese and trans-

lated to English. The answer contains links to recom-

mended products.

If no recommendation is found from our KG

triples, the designed Algorithm returns a text message

to the user informing that “This product is not com-

patible with Corsa 2004”. In our example, the KG re-

turned a not empty P

rec

set. This set is ﬁltered using I

and ranked by conversion (sales/visits). Speakers out

of stock are removed from P

rec

. In this case, our Algo-

rithm returns the text: “This product is not compatible

with Corsa 2004. However, we found some products

that ﬁt Corsa 2004”, followed by a list of links. In our

running example, there are two recommended prod-

ucts.

4 EVALUATION

This section describes our evaluation about the rec-

ommendation feature. This feature was activated for

50 customer stores of GoBots; all of them related to

the automobile domain. Subsection 4.1 describes a

quantitative analysis based on system logs whereas

Subsection 4.2 presents results of the qualitative anal-

ysis relying on collected data from sellers’ opinions.

4.1 Quantitative Evaluation

We conducted the quantitative analysis by logging the

number of recommendations produced by our solu-

tion. Among the stores that had the functionality ac-

tivated, we logged how many responded with recom-

mendations, and how many had positive results from

the sellers’ perspective.

We gradually activated the system for online

stores in the automotive sector from November 2021

to April 2022. At the end of this period, the system

had automatically responded with product recommen-

dations on thirty one (31) stores. Figure 4 shows three

graphics related to the quantitative analysis. Stores

that have not yet responded to compatibility questions

may be small stores with few products and few ques-

tions.

Of the 31 stores, there were a total of 457 re-

sponses with recommendations containing links to

other products (graphic 1 in Figure 4); 171 other re-

sponses indicated no compatibility between the cus-

tomer’s product and the store’s product; and no avail-

able product to be recommended. This resulted in a

total of 608 responses related to the recommendation

(graphic 2 in Figure 4).

The total value presented in graphic 1 in Figure 4

(457) corresponds to all responses with at least one

recommendation; those that followed the “True” path

in the decision ﬂowchart of step C in Figure 2. At

least 457 items were suggested to users who may have

purchased the recommended item. On the other hand,

the total value presented in graphic 2 in Figure 4 (628)

corresponds to all responses that included or not a rec-

ommendation; those that followed the path “True” or

“False” in the decision ﬂowchart of step C of Figure

Almost half of the recommendations from both

graphic 1 and graphic 2 (in Figure 4) are from a sin-

gle store, which is one of the biggest and with highest

sales in the e-commerce platform where this study is

situated. Graphic 3 (in Figure 4) reveals an increase

in the number of responses over the weeks analyzed.

This result was due to both the vertical scalability of

the results (number of products and recommendations

growing) and the horizontal scalability (number of ac-

tivated online stores using the recommendation fea-

ture).

The proportion between the total expressed in

Graphic 1 and in Graphic 2 in Figure 4 shows an in-

teresting metric. This is the number of recommenda-

tion responses divided by the maximum number of

possible recommendation responses. The closer to

1, the more this number demonstrates that the sellers

had products in stock compatible with the consumer’s

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

Figure 4: Sum of recommended products in three different perspectives. Graphic 1 shows that 457 responses contain at least

one recommended products; Graphic 2 presents 628 responses containing 457 responses from Graphic 1 plus 171 responses

with no products available to be recommended; Graphic 3 shows the increase of the number of recommendation answers by

several weeks of the year.

item. A value close to 0 indicates that, although the

recommendation functionality is active, no products

compatible with the consumer’s item are found.

The sellers have another way of giving feedback

on some customer responses. If the seller ﬁnds any

kind of error, whether it’s in the recommended links

or in the answer as a whole, he can mark it as incor-

rect. We identiﬁed that among the 457 recommenda-

tions, only 7 (5.87%) were marked by the sellers as

incorrect. This shows the quality of the recommenda-

tion system, correctly answering 94.13% of the ques-

tions with recommendation answers performed.

4.2 Qualitative Evaluation

Our qualitative evaluation was conducted in two

steps. We ﬁrst proposed and sent an online form to a

seller to obtain initial feedback regarding the recom-

mendation. This seller was chosen because it repre-

sents the store with more recommendation responses

and more errors pointed out by the seller. After-

wards, we conducted an online interview with this

seller to further comprehend the perception regarding

the recommendation feature in the e-commerce plat-

form. We aimed to ﬁgure out the seller’s assessment

regarding the recommendation results.

We presented and analyzed two real-world recom-

mendation answers for different products generated

by our solution regarding the recommendation feature

based on the KG. The ﬁrst product is a “Rear View

Camera for the Ecosport car 2020 model” (translated

to English language by the authors). A user asked the

following question: “Does it ﬁt in Ford Ka 2019?”

(translated to English language by the authors). Our

algorithm 1 identiﬁed the compatibility intent and

“Ford Ka 2019” as a car.

The second product was a “Screen Unlock Inter-

face without DVD player for the Audi S4 car” (trans-

lated to English language by the authors). Another

different user asked the following question: “Hi, I

have an Audi S4 2013 without DVD Player, does it

ﬁt?” (translated to English language by the authors).

Our algorithm 1 identiﬁed the compatibility intent and

“Audi S4 2013” as a car. The seller answers the ques-

tions based on these two recommendations.

Figure 5 presents the designed feedback form in

an infographic format. Our purpose was to evalu-

ate both the idea of recommending products from the

seller’s point of view and the results obtained, includ-

ing strengths and aspects that should be improved.

The form was composed of 10 questions, 9 contain-

ing multiple-alternative answers; 2 of these questions

were related to the general perception of the recom-

mendation feature (section General of Figure 5). The

Knowledge Graph-based Product Recommendations on e-Commerce Platforms

others eight questions were related to the analysis of

the results of recommendations (red and blue columns

of Figure 5). The rationale behind our evaluation

was to understand the recommendation’s qualitative

aspects in an easy-to-follow process to the seller.

As a result, the seller considered the recommen-

dation feature useful (question 1 of Figure 5). In the

second question the user had the opportunity of sug-

gesting what improvements could be accomplished in

the recommendation feature.

At this point, the seller provided his feedback

about the ﬁrst (questions 3 to 6) and second (ques-

tions 7 to 10) recommendations.

The seller answered that all products were cor-

rectly recommended (question 3) in a satisfactory way

(question 4), respecting a good order of the links

(question 5). The seller indicated the number of rec-

ommended items (3) unsatisfactory, since two links

were equal (question 6).

For the second product recommendation, the

seller informed that none of the products were cor-

rectly recommended (question 7), leading to unsatis-

factory answers regarding the recommendation’s or-

der (questions 8 to 10). In this scenario, the seller

informed that none of the products were speciﬁc to

the consumer item (Audi A4 2013).

After reaching these results, we conducted a tech-

nical exploratory analysis to identify why the ﬁrst

recommendation was successful and the second one

unsuccessful. In the ﬁrst recommendation, the user

asked about compatibility with the car, a “Ford Ka

2019”. The recommended products were all fully

compatible (t = FullCompatibility) with the car as

input from the question. The recommendation fea-

ture led the user from an incompatible product (“Rear

View Camera for the Ecosport car 2020 model”) to

products that ﬁt the car. This could have generated a

sale. Also, this ﬁtted in the consumer item and was

sold by the same store as the “Rear View Camera”.

If the recommendation feature were not activated at

the time of the question, the answer would be that the

product is not compatible, making the sale unfeasible.

The second recommendation was considered un-

successful based on the seller’s feedback. The user

(customer) asked for compatibility with his car, an

“Audi S4 2013”. They received three products that

had in their titles cars different from the consumer

item (e.g., “Screen Unlock Interface for Corolla Cross

2022”). These three products were of the “Universal

Compatibility” type (t = UniversalCompatibility),

which states that they (three different types of Screen

Unlock Interface) ﬁt in any car.

We found that the difference between the ﬁrst and

the second recommendation was the type of compat-

ibility. Customers and sellers might not ﬁnd interest-

ing the recommendation of Universal Compatibility

products.

Based on the exploratory analysis from the seller’s

answers, we found it necessary to go deeper into some

aspects. We aimed to ensure that our perception re-

garding errors and successes in the recommendations

was correct. We addressed it via an interview.

In the interview session (with the same seller that

answered the feedback form), we ﬁrst explained our

purpose with the interview and the whole process of

recommending one or more products to the seller. We

explained that the recommendation feature in the e-

commerce platform is only triggered if the product

in stock is not compatible with the consumer item

from the customer’s query. We indicated that our goal

was to determine whether the recommendation results

were correct and to which extent they could be im-

proved.

The interview was conducted in a semi-structured

way based on three questions that summarized our

doubts regarding the form answers given by the seller.

The questions were related to universal compatibility;

the number of recommended items; and ranking strat-

egy. These topics were not explored in depth in the

form step, and a direct conversation in the form of

an interview with the seller would help solving open

doubts.

Regarding the interview results, the ﬁrst relevant

point was the answer given to the question: “For your

store, is it interesting that products that have univer-

sal compatibility are recommended?”. The seller an-

swered that “universal products” are great for the rec-

ommendation strategy, but his store had not many

products of this type. Therefore, in this seller’s con-

text, there would not be a signiﬁcant difference be-

tween recommending these types of products or not.

In addition, the seller suggested that the appearance of

these products should be highlighted to the consumer

(on the webpage) if they were in the recommendation

response.

Afterwards, we aimed to further understand the

seller’s afﬁrmative (satisfactory) answer regarding the

number of recommended items. The seller mentioned

that he judged the number three as plausible because

a higher number of recommendations would be too

many. In this case, the answer would be polluted and

difﬁcult for customers to ﬁnd the ideal product.

At the end of the interview, we approached the

seller to obtain suggestions to improve the ranking

strategy. The seller reafﬁrmed that the strategy used

was sufﬁcient and that other ranking strategies, such

as ranking the products by their price, could gener-

ate unexpected effects. The seller cited, as an exam-

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development

Figure 5: Infographic containing questions about the recommendation feature and answers of a Brazilian automobile seller

where the recommendation feature is activated. The green column describes the 2 products used in the interview process and

two questions related to both recommendations. The red and blue columns are related to the questions and answers about the

ﬁrst and second recommendation, respectively.

ple, a ranking based on product price, which could be

an inefﬁcient strategy because customers in his online

store do not have preferences for more expensive and

speciﬁc products or cheaper and generic products.

5 DISCUSSION

The recommendation of products in e-commerce is a

complex task, given the diversity of users, their pref-

erences, and the applicable domains in e-commerce

platforms. Its importance is undeniable because prod-

ucts that would not be shown to the end users could

be forgotten and not sold. We understand that the

methodology presented in this research contributes to

the state-of-the-art by appearing as a novel way of

obtaining these types of recommendations. Relying

on KGs and encoded compatibilities as RDF triples,

we developed a methodology deployed in a real-world

scenario.

Our qualitative evaluation conducted showed

room for discussions and reﬁnements around univer-

sal products and other decisions, such as the number

of recommended items. These elements may vary ac-

cording to the sellers’ needs, and we understand that

the seller can customize such decisions. The impact

of these customizations as well as their implications

are subject for future studies.

Our quantitative analysis reveled the effectiveness

of our solution for automotive online stores in a great

Latin American e-commerce platform. Via our solu-

tion and initial assessments, thirty-one online stores

started to recommend products. This makes possi-

ble new sales that would have been ignored before,

given the incompatibility between consumer’s item

and the product on sale. During the four months of

our software tool monitoring, we observed an increase

in recommendations per week. The value of 5.87%

of incorrect recommendations, out of 457 recommen-

dations made, indicates an acceptable error value by

sellers.

We understand that our developed methodology

and implemented techniques are easily expandable to

other domains, as long as there is compatibility inten-

tion in texts used as input to the solution. Possible ex-

pansions of our investigation can be made by adding

knowledge about compatibility between, for example,

cell phones and their parts, clothes and people, among

other domains.

An open challenge in our study remains to the use

of data sources other than Q&A. Product descriptions,

images, videos, external resources, and compatibil-

ity tables are examples of possibilities that can en-

rich KGs with further knowledge. As a consequence,

our solution will be suited to generate more responses

containing recommendations.

Knowledge Graph-based Product Recommendations on e-Commerce Platforms

6 CONCLUSION

Having alternatives regarding products in e-

commerce plays a key role for the purchase.

However, ﬁnding and ﬁltering adequate and compat-

ible products based on customer’s natural language

questions remains a very difﬁcult task. In this article,

we proposed a novel methodology to recommend

e-commerce products based on RDF KGs. The rec-

ommendation uses questions asked by e-commerce

customers interested if their consumer item is com-

patible with the sales product. Our recommendation

technique queries a KG ﬁlled with compatibility

between consumer items and products. Our proposal

was implemented and deployed in a real-world e-

commerce platform in Latin America. We found that

our solution is suitable for recommending relevant

products in the e-commerce platform. Future work

involves the study of novel methods to compare

available products on e-commerce platforms. This

work used textual properties, such as categories,

store, and domain. The recommendation algorithm

could further beneﬁt from comparing NL texts and

images, for instance. In addition, based on the

evaluation results, we plan to further assess the use

of products that are universally compatible within the

recommendation feature.

ACKNOWLEDGEMENTS

This study was ﬁnanced by the National Council for

Scientiﬁc and Technological Development - Brazil

(CNPq) process number 140213/2021-0.

REFERENCES

Covington, P., Adams, J., and Sargin, E. (2016). Deep

neural networks for youtube recommendations. In

Proceedings of the 10th ACM Conference on Rec-

ommender Systems, RecSys ’16, page 191–198, New

York, NY, USA. Association for Computing Machin-

ery.

Dwivedi, R., Anand, A., Johri, P., Banerji, A., and Gaur,

N. (2020). Product based recommendation system on

amazon data. International Journal of Creative Re-

search Thoughts - IJCRT.

Guo, Q., Zhuang, F., Qin, C., Zhu, H., Xie, X., Xiong, H.,

and He, Q. (2020). A survey on knowledge graph-

based recommender systems. IEEE Transactions on

Knowledge & Data Engineering, (01):1–1.

Isinkaye, F. O., Folajimi, Y., and Ojokoh, B. A. (2015). Rec-

ommendation systems: Principles, methods and eval-

uation. Egyptian informatics journal, 16(3):261–273.

Lee, D. and Hosanagar, K. (2014). Impact of recommender

systems on sales volume and diversity. In Myers,

M. D. and Straub, D. W., editors, Proceedings of the

International Conference on Information Systems -

Building a Better World through Information Systems,

ICIS 2014, Auckland, New Zealand, December 14-17,

2014. Association for Information Systems.

Linden, G., Smith, B., and York, J. (2003). Amazon. com

recommendations: Item-to-item collaborative ﬁlter-

ing. IEEE Internet computing, 7(1):76–80.

Luo, X., Liu, L., Yang, Y., Bo, L., Cao, Y., Wu, J., Li, Q.,

Yang, K., and Zhu, K. Q. (2020). Alicoco: Alibaba

e-commerce cognitive concept net. In Proceedings of

the 2020 ACM SIGMOD International Conference on

Management of Data, pages 313–327.

Sant’Anna, D. T., Caus, R. O., dos Santos Ramos, L.,

Hochgreb, V., and dos Reis, J. C. (2020). Generat-

ing knowledge graphs from unstructured texts: Expe-

riences in the e-commerce ﬁeld for question answer-

ing. In Advances in Semantics and Linked Data: Joint

Workshop Proceedings from ISWC 2020, pages 56–

71.

Shao, B., Li, X., and Bian, G. (2021). A survey of research

hotspots and frontier trends of recommendation sys-

tems from the perspective of knowledge graph. Expert

Systems with Applications, 165:113764.

Wu, P., Zhang, X., and Feng, Z. (2019). A survey of ques-

tion answering over knowledge base. In Zhu, X., Qin,

B., Zhu, X., Liu, M., and Qian, L., editors, Knowledge

Graph and Semantic Computing: Knowledge Com-

puting and Language Understanding, pages 86–97,

Singapore. Springer Singapore.

KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development