Query and Product Suggestion for Price Comparison Search Engines

based on Query-product Click-through Bipartite Graphs

Lucia Noce, Ignazio Gallo and Alessandro Zamberletti

University of Insubria, Department of Theoretical and Applied Science,

Via Mazzini, 5, 21100 Varese, Italy

Keywords:

Query Suggestion, Product Suggestion, Price Comparison Engines.

Abstract:

Query suggestion is a technique for generating alternative queries to facilitate information seeking, and has

become a needful feature that commercial search engines provide to web users. In this paper, we focus on

query suggestion for price comparison search engines. In this speciﬁc domain, suggestions provided to web

users need to be properly generated taking into account whether both the searched and the suggested products

are still available for sale. To this end, we propose a novel approach based on a slightly variant of classical

query-URL graphs: the query-product click through bipartite graph. Such graph is built using information

extracted both from search engine logs and speciﬁc domain features such as categories and products popular-

ities. Information collected from the query-product graph can be used to suggest not only related queries but

also related products. The proposed model was tested on several challenging datasets, and also compared with

a recent competing query suggestion approach speciﬁcally designed for price comparison engines. Our solu-

tion outperforms the competing approach, achieving higher results both in terms of relevance of the provided

suggestions and coverage rates on top-8 suggestions.

1 INTRODUCTION

Query suggestion plays an important role in helping

users to ﬁnd what they are looking for when query-

ing web search engines. It is a particularly interest-

ing research topic because the usability, popularity

and success of web search engines are strongly re-

lated to the effectivenessof the helping tools provided

to users while performing textual queries: the larger

the amount of users that issue queries leading to opti-

mal/desirable results, the higher is the usability of the

web search engine and consequently the larger is the

amount of users that will keep using that web search

engine for other future queries.

Although many effective and sophisticated query

suggestion algorithms that have been proposed in lit-

erature are currently employed by web search engines

to improve user search experience (Kato et al., 2013),

precisely understanding in a limited amount of time

the needs of service users from textual queries is still

a challenging task that has yet to ﬁnd an optimal so-

lution. In fact, achieving satisfying query suggestion

accuracies for textual queries is a non-trivial task be-

cause most user-made textual queries are typically

very short and contain ambiguous words, and it is

therefore difﬁcult to provide good query suggestions

without using complex pipelines.

Most query suggestion algorithms in literature fo-

cus on improving user search experience by suggest-

ing related queries from a given textual query, try-

ing to guide users in ﬁnding what they are looking

for. The list of suggestions is usually created by

mining related queries from web search engine logs

and session information, trying to gather previous

users knowledge. Depending on the chosen approach,

query suggestion techniques can be grouped into two

categories (Meng, 2014): session-based methods and

click-through-based methods. The ﬁrst approach uses

the consecutiveness of the queries to model user be-

havior within each user session, while the second

one focuses on the relationship between the sub-

mitted queries and the URLs clicked by the user.

Both session-based methods and click-through-based

methods are being used by many commercial web

search engine, e.g. Google, Ask and Yahoo!, to pro-

vide correlated queries to improve usability.

Despite most of the query suggestion works in lit-

erature were developed for web search engines, query

suggestion is also important for e-commerce plat-

forms and price comparison search engines (Hasan

Noce, L., Gallo, I. and Zamberletti, A.

Query and Product Suggestion for Price Comparison Search Engines based on Query-product Click-through Bipartite Graphs.

In Proceedings of the 12th International Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 1, pages 17-24

ISBN: 978-989-758-186-1

et al., 2011). In this work we present a click-through-

based query suggestion system speciﬁcally designed

for price comparison websites (ShoppyDoo, 2015). It

provides query suggestions on the basis of all the spe-

ciﬁc product information available in this context, e.g.

the category of the product that the user is browsing

while issuing the query and the clicked product. To

this end, instead of using the URLs clicked by users,

our click-though based method exploits the informa-

tion from clicked products.

Query suggestion in price comparison websites is

an even more challenging task because, for exam-

ple, it may be disadvantageous to suggest a query for

which there are no available product offers.

2 RELATED WORKS

Query suggestion is an interesting research topic

widely investigated in researches on Internet search

domain (Kim and Croft, 2014; Song et al., 2012;

Jiang et al., 2014). Algorithms for query suggestion

can be classiﬁed as either session-based (Zhang and

Nasraoui, 2008; Boldi et al., 2009; Lau and Horvitz,

1999; Zanon et al., 2012; Ozmutlu and Spink, 2003)

or click-through-based (Mei et al., 2008; Ma et al.,

2010; Song and He, 2010).

Session-based methods assume that users submit

queries in a sequential manner. The basic assump-

tion is that when a badly written textual query is ex-

ecuted by a user, there will always be a correct ver-

sion of the wrong textual query previously executed

by the same user that follows the wrong one. The

sequence of queries is usually obtained by exploiting

user sessions, and it is used to extract the user’s inten-

tions. Among session-based query suggestion meth-

ods, the works that are closely related to our are the

ones of (Lau and Horvitz, 1999; Ozmutlu and Spink,

2003; Boldi et al., 2009; Zanon et al., 2012).

In their work about query reﬁnement Lau et

al. (Lau and Horvitz, 1999) assert that, after a failed

search, most users reﬁne their original query either

by adding more details or by executing a new one.

Another study by Ozmutlu and Spink (Ozmutlu and

Spink, 2003) demonstrates that in 88.6% of cases a

query session relates to a just a single topic, support-

ing the thesis that query sessions are useful to extract

information that can be used for query suggestion pur-

poses.

Following this approach, Boldi et al. (Boldi et al.,

2009) propose a system that uses query-ﬂow graphs

built from query session logs. In the query-ﬂow

graph, there exists an edge from a query q

to a query

if and only if both queries belong to the same ses-

sion. Edges and vertexes can be weighted. These

weights represent the frequency of a query and the

popularity of a transition for a vertex and for an edge

respectively. Edges are also labelled with some ex-

tra information describing the nature of the transition

(e.g. specialization, probability that user moves from

to q

, etc.). In their experimental evaluation, Boldi

et al. show that adding weights and labels to query-

ﬂow graphs allows the creation of a session-based

query suggestion algorithm which achieves the same

results of click-through-based models.

A session-based query suggestion method specif-

ically designed for price comparison engines has

been proposed by Zanon et al. (Zanon et al., 2012).

The model performs query suggestion system using

the same query-ﬂow graphs proposed by Boldi et

al. (Boldi et al., 2009). Zanon et al. adapt the model

of Boldi et al. to their context exploiting the fact that

in price comparison engines offers are grouped into

categories. Their model reach a quality of 70% and a

coverage of 37% for top-8 suggestions.

Click-through-based methods exploit the informa-

tion of the URL that a user clicks after having sub-

mitted a query. These category of approaches assume

that two queries are similar if they share a consistent

number of clicked URLs.

Baeza-Yates et al. (Baeza-Yates et al., 2004) ex-

ploit click-through data in their query recommenda-

tion method. The presented methodology is based

on a clustering process over data extracted from the

query log, where suggested queries are proposed ac-

cording to a rank score evaluated in terms of similarity

and support of the queries.

Following Baeza-Yates et al. work, a query sug-

gestion method for e-commerce was proposed by

Hasan et al. (Hasan et al., 2011). Authors discuss

about the challenges related to their context and un-

derline that, due to the nature of the data belong-

ing to e-commerce websites, it could be difﬁcult to

adapt session-based query suggestion approaches to

e-commerce systems. Instead, a common approach

in this ﬁeld is to use graph representation for under-

lining queries and URLs relationship through a click-

through bipartite graph. Click-through bipartite, rep-

resents an implicit judgment given by users of the re-

lationships between queries. It has been proven that

this type of graph is a very precious source of infor-

mation for measuring similarities among its compo-

nents (Wu et al., 2013).

Many click-through-based approaches exploit bi-

partite graph, e.g. Ma et al. (Ma et al., 2010) proposed

a ranking method that is able to diversify query sug-

gestion results, employing Markov random walk pro-

cess and hitting time analysis; Song et al. use bipar-

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

tite graph to analyze both clicked URLs and skipped

URLs to recommend related queries (Song and He,

2010); Cao et al. (Cao et al., 2008) exploit click-

through data and group similar queries into concepts,

providing query suggestions based on these concepts.

3 PROPOSED METHOD

Query suggestion in price comparison websites is

very similar to the query suggestion in e-commerce

websites, because in both scenarios users search for

products they want to buy and therefore it is reason-

able to say that they issue to the two different search

engine the same type of queries. Following the dis-

cussion of Hasan et al. (Hasan et al., 2011) about e-

commerce websites, the proposed method is a click-

through-based query suggestion approach.

Our model exploits the two main entities of the

website for which it has been designed for (Shoppy-

Doo, 2015): products and categories. These concepts

are not restricted to our context but are used in many

other major e-commerce and search engine websites.

More in details, the result of a generic query in our

context is a list of commercial products sold by sev-

eral retailers and sorted by descending price. Each

item of the list is an offer. All the offers belong

to a speciﬁc category, according to the type of the

objects searched by the user. Offers may also be

linked to products. Products represent aggregations

of items/offers, and all the offers inherent to a speciﬁc

item are collected under the related product.

In general, a click-through-based query sugges-

tion algorithm is composed of two phases: (i) an of-

ﬂine module which manages the data and builds the

model, (ii) and an online procedure which supplies

the query suggestions. Our solution extracts informa-

tion from web server logs. Starting from those logs,

the ofﬂine phase of our method can be summarized in

the following steps:

• Data Extraction: is the pre-processing phase, the

web server log is parsed. Queries and related in-

formation are extracted, cleaned and normalized;

• Session creation: the sessions are created as asso-

ciations between queries and clicked products;

• Building phase: creation of the click-through bi-

partite graph by exploiting the notion of product;

On the other hand, the subsequent online phase

gathers query suggestions from the model and returns

a ﬁxed number of ranked related queries and products.

In Figure 1, both the ofﬂine and the online phases

of the proposed approach are visually summarized. In

the following sections all their major details are re-

ported.

Figure 1: Visual summarization of the proposed pipeline.

From top to bottom: starting from raw log data, all the in-

formation needed to determine query sessions are extracted;

the query-product bipartite graph is built and suggestions

are provided by exploiting both the product categories and

the product popularities extra-ﬁelds.

3.1 Data Extraction

The ﬁrst step consists of the extraction of data from

the web server log. Those logs contains rough data

that has to be cleaned before being used for query sug-

gestion.

Different cleaning stages are performed to gather

only log lines containing the information we want to

exploit. First we performed bot ﬁltering to retrieve

only log search lines that lead to a user click. Within

a log line we extract only the ﬁelds of interest: the tex-

tual query, the id of the category for which the query

has been performed, and the clicked product. Syn-

onym mapping for queries like “mtb” and “mountain

Query and Product Suggestion for Price Comparison Search Engines based on Query-product Click-through Bipartite Graphs

Figure 2: This visual example shows the usage of the query-product bipartite graph. Some details, such as the popularity

score and the number of clicks, are omitted to make the image more readable. Assuming that the input query is equal to q

and that N (number of suggestions) is equal to 3, we determine p

, p

and p

as the 3 most popular products related to the

input query q

, and we gather from them the three most clicked related queries q

, q

and q

bike” and spelling correction for cases such as “dy-

landog” corrects into “dylan dog” are also performed.

We also use stemming to detect equivalent queries,

e.g. “woman bike” is equivalent to “women bikes”,

“samsung galaxy S6” represents the same query as

“S6 samsung galaxy”. Since the website is in italian,

all the reported examples have been translated for bet-

ter understanding.

3.2 Session Creation

For building the model, we take into account only tex-

tual queries extracted from web server logs. A query

is the list of terms that a user types to search for the

item he needs, the same query may be issued to the

search engine several times, each submission creates

a query session.

We followed the approach that has been proposed

by Wen et al. (Wen et al., 2001) and used by Baeza-

Yates et al. (Baeza-Yates et al., 2004) which considers

a simple notion of query session that is composed of

a query and the URLs clicked within its query results.

In our method instead of URLs we collected the prod-

ucts returned by the query, as follows:

QuerySession = (searchedText, clickedProducts)

The reason why we adopt a different notion of

query session lies our application context. In price

comparison engines, each time a query is submitted,

a list of related offers is shown. However, each offer

may be available for a limited amount of time, there-

fore it would be senseless to gather the direct offers

URLs which may have been already destroyed. Cre-

ating query suggestions dealing to unavailable offers

is improper and unattractive for users, and may cause

a loss of clients for the website.

The taxonomy of our price comparison website

allows us to exploit the notion of product. Not ev-

ery offer is related to a product, but statistical studies

made by the website administrator shows that the ma-

jority (97%) of the offers dealing to a user click are

related to a product. Starting from this assertion we

consider only offers related to products and sessions

storing products related to the submitted queries. This

not only allows us to provide a more reliable query

suggestion system, but also permits to supply product

suggestions to users.

3.3 Building Phase

We build the proposed model using the notion of

query session previously described. This way we cre-

ate a different version of the well-known query-URL

bipartite graph in which, instead of URLs, we use

clicked products.

A query-product bipartite graph consists of two

disjoint sets of nodes corresponding to queries and

products respectively. Intuitively, in a click-through

graph vertexes on one side correspond to queries,

while vertexes on the other side represents products.

An edge connecting a query q to a product p exists

whenever p was clicked starting from q.

A visual example of such query-product bipartite

graph is given in Figure 2. The set of nodes in the

left part are queries and the ones in the right part are

products. An edge between a query q and an URL u

indicates a user click of u when issuing q (for simplic-

ity click numbers are omitted from the graph).

The query-product bipartite graph gathers a large

amount of potential information that can be used for

query suggestion, query clustering, query reformula-

tion and so on.

In our method we consider two kind of graphs

depending on whether a user uses a category when

searching for a product or not. In our price com-

parison website, users are allowed to submit queries

within a speciﬁed category, and they expect to receive

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

product results from that same category. Showing

query suggestions outside the selected category may

be annoying for the users; as such so build two dif-

ferent kind of graphs. The ﬁrst one does not exploit

the concept of category and therefore it can be used to

provide query suggestion results from all the website

categories. On the other hand, query suggestion re-

sults from the second graph are restricted to a speciﬁc

category selected by the user. By query suggesion re-

sults we intend both queries and products.

A ﬁnal pruning phase is necessary to avoid the

suggestion of queries that may have a negative impact

on the user experience. We start removing products

that are only related to one query, and then we remove

queries that are only connected to one product.

Both the two versions of the query-product bipar-

tite graph were tested and evaluated both in terms of

time needed to complete their building phase, and in

term of accuracy for the provided query suggestion

results. All the experiments and results are reported

and discussed in Section 4.

3.4 Online Query Suggestion

The online phase of our method exploits two basic

concepts related to the taxonomy of the price com-

parison engine we used in this work: categories and

popularity. As previously described, for query sug-

gestion, we take into account only offers that belong

to a product.

For each product the website administrator has

provided us a numeric value which indicates the pop-

ularity of an offer. This value is computed considering

both the whole server trafﬁc and the available offers in

a certain period of time, and it is periodically updated.

Given a user query q, if it belongs to the query

set of the bipartite graph, we automatically obtain the

list of the related queries and of the related products.

The retrieved product list is then sorted by ascending

popularity value, and the top k products are collected.

Starting from this list of top k products, another list of

related queries is produced and sorted by the number

of clicks stored in the product-query bipartite graph

edges. The top k retrieved queries are selected and

provided as results.

On the other hand if a query q does not belong to

the query-product bipartite graph we use the similar-

ity measure described by Zanon et al. (Zanon et al.,

2012) to retrieve the more similar query q

′

among the

bipartite graph. This measure exploits the concept of

categories and the Jaccard similarity coefﬁcient pro-

posed by Tan et al. (Tan et al., 2005). Once q

′

is found

the system acts as previously described.

4 EXPERIMENTS

We conducted several experiments to measure the ef-

fectiveness of our proposed query and product sug-

gestion method. Both query and product suggestions

were evaluated with and without using category infor-

mation.

To build our query-product bipartite graph we

have used three months of ShooppyDoo user query

logs. More in details, we have collected a total of

1147990 queries, 1096112 clicked products and we

have built a ﬁnal version graph containing 79025

unique queries, 283188 unique products. The total

amount of available products is 295883, so with three

months of data log we almost coveredall the products.

In Table 2 some more details for the graph build-

ing phase are provided. We report both the time

needed by our model to complete the graph building

phase using one, two and three months of query logs;

and the details of the product-query bipartite graph in

terms of number of queries and products after and be-

fore the pruning phase. All the experiments have been

performed on a Intel Core i5 @ 3.4 GHz with 12 GB

of RAM; the software has been developed using Java

for the pre-processing phase, and MATLAB for the

building and experimental phases.

4.1 Dataset

We create a test dataset following the manner adopted

by Song et al. (Song and He, 2010; Song et al., 2012).

Three sets of 500 queries each are randomly se-

lected from the query log ﬁle. The ﬁrst contains 500

queries not related to any category, the second con-

tains 500 queries related to one of the 569 categories

of the websites (suggestions also belong to the same

category), while the third contains a balanced equal

number of queries related to 5 selected categories.

For each sampled query our algorithm generates top-8

suggestions for both queries and products.

Three judges were asked to evaluate the quality of

our query suggestion model. In order to avoid a po-

sitional bias, we mix the provided suggestions during

the evaluation phase. More in details, we have pro-

vided them the initial query along with 8 queries sug-

gestions and 8 product suggestions. We have also pro-

vided some extra information, such as the category for

which the query was originally executed to let them

better understand the context and the meaning of the

query. Judges could choose between three different

values: relevant, irrelevant or no opinion (difﬁcult to

decide) for both types of suggestions.

For each query we have obtained two ﬁnal rates,

one for query suggestion and one for product sugges-

Query and Product Suggestion for Price Comparison Search Engines based on Query-product Click-through Bipartite Graphs

Table 1: The ﬁve different categories adopted during the experimental phase have substantially different characteristics: the

“Food Supplements” category is the smallest one and less products are associated to it; the “Fridge category” is the most

popular one. The cardinality of each category deeply affects the quality of the query/product suggestions provided by our

model, as reported in Table 3 and 4.

Category Tot # products # products with offers # offers

Mobile phone 5040 847 8110

Oven 4735 1293 13328

Fridge 8166 1908 22503

Food supplements 71 68 485

Tires 1896 1635 12900

Table 2: Building phase details. For each query log size

(1, 2 or 3 months), the ﬁrst line represents contains the size

of the pruned bipartite-graph, while the second line reports

the initial numbers of queries and products. Bold values

denoted the ﬁnal model used during the experimental phase.

# months Time (s) # queries # products

3 37.56

79025 283188

156434 283876

2 18.25

49855 141765

98654 205465

1 6.61

48926 48930

675341 62567

Table 3: Query suggestion coverage rates calculated on all

the 7 test dataset. The model provides better suggestions for

the more populated categories (e.g. Oven) than it does for

the less populated one (e.g. Tires). Also, the exploitation of

the category popularity value enables to reach higher results

within the “categorized” dataset compared to the “uncatego-

rized” one.

Dataset 3 sugg. 5 sugg. 8 sugg.

Uncategorized 73.36 % 63.48 % 56.40 %

Categorized 79.53 % 78.94 % 75.53 %

Mobile phone 95.21 % 92.11 % 85.32 %

Oven 92.45 % 84.90 % 81.12 %

Fridge 94.17 % 90.07 % 89.75 %

Food supplements 69.09 % 62.04 % 34.61 %

Tires 32.56 % 23.11 % 11.56 %

tion. Those rates have been calculated using the ma-

jority vote among the three judges evaluations.

4.2 Evaluation Measure

We evaluate the quality of the two types of sugges-

tions provided by the model using Precision at rank N

(P@N) and Mean Average Precision (MAP).

The precision at rank N, P(N), for a speciﬁc query is

deﬁned as the percentage of relevant queries:

P(N) =

# relevant queries

(1)

P@N is deﬁned as the aggregated precision for all

queries:

Table 4: Product suggestion coverage rates. The suggested

products reach an high coverage rate among all the different

test datasets. The lowest values are achieved applying the

model to the less populated category, and the category val-

ues exploited on the “categorized” dataset helps the model

in reaching the highest results.

Dataset 3 sugg. 5 sugg. 8 sugg.

Uncategorized 73,68 % 63,15% 44.68 %

Categorized 89,42% 85,62% 81,52%

Mobile phone 72,10% 63,22% 54,78 %

Oven 79,24% 75,47 % 70,58 %

Fridge 83,13% 78,01% 76,31 %

Food supplements 68,03 % 51,89 % 38,46 %

Tires 50,07 % 27,34 % 8,45 %

P@N =

∑

P(N)

# total queries

(2)

MAP is the average of Average Precision for all

queries. For each query q

AverageP(q

) =

∑

P( j) · rel( j)

# relevant queries

(3)

where the precision at rank j, P( j) is deﬁned as in

equation 1, and rel( j) is equal to 1 when the item at

rank j is relevant, 0 otherwise. The mean average

precision for a set of queries Q is the mean of the

average precision scores for each query q

MAP =

∑

i=1

AverageP(q

)

(4)

We also adopt a coverage metric to indicate for

how many input queries the algorithm returns at least

a minimum amount of query/product suggestions. We

calculated coverage rate considering a minimum of 3,

5 and 8 queries.

4.3 Results

Evaluation metrics are calculated on three different

type of datasets built as described in Section 4.1; First

dataset, called ”uncategorized“, contains queries not

related to any category; second dataset, called ”cat-

egorized“, contains queries and query results within

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

Table 5: Overall scores for the query suggestions task expressed in terms of MAP and P@N, considering the top-3, top-5 and

top-8 suggestions.

Dataset MAP P@1 P@3 P@5 P@8

Uncategorized 0.76 0.92 0.86 0.78 0.72

Categorized 0.81 0.94 0.89 0.81 0.75

Mobile phone 0.81 0.98 0.93 0.80 0.75

Oven 0.79 0.95 0.83 0.79 0.72

Fridge 0.80 0.98 0.91 0.83 0.77

Food supplements 0.59 0.82 0.64 0.59 0.30

Tires 0.34 0.70 0.29 0.21 0.09

Table 6: Overall scores for the product suggestions task expressed in terms of MAP and P@N, considering the top-3, top-5

and top-8 suggestions.

Dataset MAP P@1 P@3 P@5 P@8

Uncategorized 0.79 0.97 0.82 0.75 0.68

Categorized 0.82 0.95 0.87 0.80 0.71

Mobile phone 0.76 0.98 0.83 0.75 0.63

Oven 0.71 0.94 0.69 0.71 0.62

Fridge 0.80 0.98 0.79 0.71 0.64

Food supplements 0.59 0.81 0.64 0.49 0.34

Tires 0.39 0.76 0.41 0.25 0.07

a randomly chosen category; (iii) third dataset is

formed by extracting queries from we 5 different cat-

egories: mobile phones, ovens, fridges, food supple-

ments and tires. For this third dataset we refer to the

associated set of queries using their category name.

The category datasets are extracted from cate-

gories that present substantially different characteris-

tics and cardinalities. Table 1 shows, for each cate-

gory, the total number of products belonging to the

category, the number of products which have some

matched offers, and the number of offers that are

matched to products. Note that we did not chose only

the most popular categories, but we also took into

account categories having a low number of products

(such as food supplements) in order to evaluate how

the model deals with those less populated categories.

Coverage rate values reported in Tables 3 and 4

reﬂect the differences among the categories: the more

populated categories reach higher results for both

query and product suggestions. The exploitation of

the category rank values in the ”categorized“ dataset

enables the model to propose more results for both

query and product suggestions, where the highest

coverage values are achieved for the product sugges-

tion task.

Tables 5 and 6 summarize the obtained results for

the query suggestion model and the product sugges-

tion model respectively. Results are shown in term of

MAP and P@N, while varying N from 1 to 8.

From our experimental evaluation results we can

conclude that our model proposes reliable suggestions

on the top-8 result not only using the ”categorized“

dataset (75% query, 71% product), but also using the

”uncategorized“ dataset (72% query, 68% product).

When referring to top-5 results precisions are around

80% for both datasets. Among categories, accord-

ing to the coverage measure, the higher results are

achieved for the more populated class.

Due to the similarity of the context of our work

and the one from Zanon et al. (Zanon et al., 2012)

and despite the fact that we used a different dataset,

we propose a weak but but still signiﬁcant compari-

son between our results and those achieved by their

model. Unlike all the other works in literature, Zanon

et al. (Zanon et al., 2012) evaluate their query sugges-

tion model on two price comparison engines, reaching

a coverage rate equals to 37% on the top-8 suggestion

and an overall quality equals to 70%.

Our model overcomes those results, achieving sig-

niﬁcantly higher coverage rate: 75% on the ”un-

categorized“ dataset and 56% on the ”categorized“

dataset. The quality of our query/product suggestion

method is also higher than theirs for top-8 results.

This conﬁrms that a click-through based model better

ﬁts query suggestions task in price comparison and

more in general in e-commerce website, as asserted

in (Hasan et al., 2011).

5 CONCLUSIONS

In this paper we have presented a novel click-through-

Query and Product Suggestion for Price Comparison Search Engines based on Query-product Click-through Bipartite Graphs

based query suggestion model speciﬁcally designed

for price comparison websites.

Unlike most of the other generic query sugges-

tion methods proposed in literature, to reach satis-

fying query/product suggestion results our approach

takes advantage of most of the relevant information

available for this category of website: product offers

and product categories.

Similarly to other methods, these information are

used to build a click-through bipartite graph. How-

ever, instead of using exploiting the association be-

tween queries and URLs clicked by users as in

traditional click-through-based query suggestion ap-

proaches, our click-through bipartite graph stores as-

sociations between user queries and clicked products.

Since in most price comparison websites product of-

fers are clustered into products, our customized click-

through bipartite graph allows our model to provide

query and product suggestions based on products and

not on direct URLs of product offers which may not

be available anymore at the time that the suggestions

is generated.

We evaluated our system both in terms of cover-

age rates and quality of the results for both types of

generated suggestions (query and products), reaching

high precision values, satisfying coverage rates, out-

performing also the results of a competing recently

published approach also speciﬁcally designed for the

same query/product suggestion tasks for price com-

parison websites.

REFERENCES

Baeza-Yates, R., Hurtado, C., and Mendoza, M. (2004).

Query recommendation using query logs in search en-

gines. In Proceedings of the 2004 International Con-

ference on Current Trends in Database Technology.

Boldi, P., Bonchi, F., Castillo, C., Donato, D., and Vigna, S.

(2009). Query suggestions using query-ﬂow graphs.

In Proceedings of the 2009 Workshop on Web Search

Click Data.

Cao, H., Jiang, D., Pei, J., He, Q., Liao, Z., Chen, E., and Li,

H. (2008). Context-aware query suggestion by min-

ing click-through and session data. In Proceedings of

the 14th ACM SIGKDD International Conference on

Knowledge Discovery and Data Mining.

Hasan, M. A., Parikh, N., Singh, G., and Sundaresan, N.

(2011). Query suggestion for e-commerce sites. In

Proceedings of the Fourth ACM International Confer-

ence on Web Search and Data Mining.

Jiang, D., Leung, K. W., Vosecky, J., and Ng, W. (2014).

Personalized query suggestion with diversity aware-

ness. In IEEE 30th International Conference on Data

Engineering, Chicago, ICDE 2014, IL, USA, March

31 - April 4, 2014.

Kato, M. P., Sakai, T., and Tanaka, K. (2013). When do

people use query suggestion? a query suggestion log

analysis. Inf. Retr., 16(6):725–746.

Kim, Y. and Croft, W. B. (2014). Diversifying query sug-

gestions based on query documents. In Proceedings of

the 37th International ACM SIGIR Conference on Re-

search & Development in Information Retrieval.

Lau, T. and Horvitz, E. (1999). Patterns of search: Analyz-

ing and modeling web query reﬁnement. In Proceed-

ings of the Seventh International Conference on User

Modeling.

Ma, H., Lyu, M. R., and King, I. (2010). Diversifying query

suggestion results. In Proceedings of the Twenty-

Fourth AAAI Conference on Artiﬁcial Intelligence.

Mei, Q., Zhou, D., and Church, K. (2008). Query sug-

gestion using hitting time. In Proceedings of the

17th ACM Conference on Information and Knowledge

Management.

Meng, L. (2014). A survey on query suggestion. Inter-

national Journal of Hybrid Information Technology,

7(6):43–56.

Ozmutlu, S., O. H. C. and Spink, A. (2003). Multitasking

web searching and implications for design. In Pro-

ceedings of the American Society for Information Sci-

ence and Technology.

ShoppyDoo (2015). http://www.shoppydoo.it/.

Song, Y. and He, L.-w. (2010). Optimal rare query sugges-

tion with implicit user feedback. In Proceedings of the

19th International Conference on World Wide Web.

Song, Y., Zhou, D., and He, L.-w. (2012). Query sugges-

tion by constructing term-transition graphs. In Pro-

ceedings of the Fifth ACM International Conference

on Web Search and Data Mining.

Tan, P.-N., Steinbach, M., and Kumar, V. (2005). Introduc-

tion to Data Mining, (First Edition). Addison-Wesley

Longman Publishing Co., Inc.

Wen, J.-R., Nie, J.-Y., and Zhang, H.-J. (2001). Clustering

user queries of a search engine. In Proceedings of the

10th International Conference on World Wide Web.

Wu, W., Li, H., and Xu, J. (2013). Learning query and doc-

ument similarities from click-through bipartite graph

with metadata. In Proceedings of the Sixth ACM Inter-

national Conference on Web Search and Data Mining.

Zanon, R., Albertini, S., Carullo, M., and Gallo, I. (2012). A

new query suggestion algorithm for taxonomy-based

search engines. In Proceedings of the International

Conference on Knowledge Discovery and Information

Retrieval.

Zhang, Z. and Nasraoui, O. (2008). Mining search engine

query logs for social ﬁltering-based query recommen-

dation. Appl. Soft Comput., 8(4):1326–1334.

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies