Personalized Web Search via Query Expansion based on User’s Local

Hierarchically-Organized Files

∗

Gianluca Moro

, Roberto Pasolini

and Claudio Sartori

Department of Computer Science and Engineering, University of Bologna, Via Venezia 52, Cesena (FC), Italy

Department of Computer Science and Engineering, University of Bologna, Viale Risorgimento 2, Bologna (BO), Italy

Keywords:

Information Retrieval, Personalized Search, Query Expansion, Local Files, Search Engine.

Abstract:

Users of Web search engines generally express information needs with short and ambiguous queries, leading

to irrelevant results. Personalized search methods improve users’ experience by automatically reformulating

queries before sending them to the search engine or rearranging received results, according to their speciﬁc

interests. A user proﬁle is often built from previous queries, clicked results or in general from the user’s

browsing history; different topics must be distinguished in order to obtain an accurate proﬁle. It is quite

common that a set of user ﬁles, locally stored in sub-directory, are organized by the user into a coherent

taxonomy corresponding to own topics of interest, but only a few methods leverage on this potentially useful

source of knowledge. We propose a novel method where a user proﬁle is built from those ﬁles, speciﬁcally

considering their consistent arrangement in directories. A bag of keywords is extracted for each directory

from text documents within it. We can infer the topic of each query and expand it by adding the corresponding

keywords, in order to obtain a more targeted formulation. Experiments are carried out using benchmark data

through a repeatable systematic process, in order to evaluate objectively how much our method can improve

relevance of query results when applied upon a third-party search engine.

1 INTRODUCTION

Millions of people daily use Web search engines to

ﬁnd Web pages they are looking for. Every search

starts from an information need expressed by the user

as a textual query. Due to haste and laziness, users ex-

press complex information needs with few keywords,

often resulting in an ambiguous query yielding many

irrelevant results (Jansen et al., 2000; Carpineto and

Romano, 2012). In a typical example, a user inter-

ested in apple-based recipes may simply search for

“apple” and to have then to set apart resulting pages

which actually deal with Apple computers.

We can disambiguate the query, in order to better

express the information need and get proper results,

by enriching the query with context. The query could

for example include the intended topic of discussion,

e.g. for “apple” it could be “fruits” or “computers”.

As every user often has a speciﬁc set of topics of inter-

est, these can be used to properly disambiguate simple

∗

This work was partially supported by the project “Tore-

ador”, funded by the European Union’s Horizon 2020 re-

search and innovation programme under grant agreement

No 688797.

queries without requiring him or her to explicitly pro-

vide a context: a cooking-enthusiast user could effort-

lessly query for “apple” to only retrieve pages about

apple-based recipes.

Personalized search provides a customized expe-

rience to the speciﬁc user or for the speciﬁc task being

carried on, leveraging contextual information. Per-

sonalization is usually wrapped around an existing

search engine using one or both of two approaches:

queries can be expanded or otherwise rewritten before

sending them to the search engine, while results can

be ﬁltered or reordered to move the most pertinent on

top (Pitkow et al., 2002).

Personalized search users are described by proﬁles

usually indicating topics or concepts of their interest,

such as “travels”, “acoustic guitar”, “videogames”

and alike. Users can be proﬁled by monitoring their

activity unobtrusively, without their explicit feedback.

It is common to leverage the history of previous

searches, including issued queries and visited results,

while other solutions mine knowledge from the whole

history of browsed pages (Ghorab et al., 2013).

A less common approach is to extract information

from documents stored locally by the user on his or

Moro G., Pasolini R. and Sartori C.

Personalized Web Search via Query Expansion based on Userâ

Zs Local Hierarchically-Organized Files.

DOI: 10.5220/0006486401570164

In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KDIR 2017), pages 157-164

ISBN: 978-989-758-271-4

her own PC or device (Chirita et al., 2007). Users

often store many text ﬁles, either self-produced or

obtained externally, representing their recurring in-

terests. These ﬁles are stored within an arbitrarily

structured hierarchy of directories, in order to eas-

ily browse the collection and ﬁnd any needed docu-

ment. We claim that these directories naturally repre-

sent speciﬁc topics of interest for the user and that

their hierarchical structure approximates a coherent

taxonomy of these topics.

We propose a personalized search solution lever-

aging these hierarchically organized local documents.

Text ﬁles are analyzed in order to extract recurring

keywords for each directory, deemed to represent a

speciﬁc topic of interest. Whenever a search engine is

used, keywords of its inferred topic are added to the

issued query beforehand, in order to disambiguate it

according to user’s interests. We use well known in-

formation retrieval (hence IR) techniques to extract

keywords, weigh their relevance and determine the

most likely topic for each query.

Compared to similar solutions, our method re-

quires neither a predeﬁned taxonomy of concepts nor

the classiﬁcation of data into such taxonomy, which

can be complex and error-prone. Instead, our method

exploits the work which users already do to organize

personal documents, in order to build taxonomies of

topics upon their own needs: no computational effort

is required to distinguish topics and to classify doc-

uments within them. The personalization system just

extracts relevant keywords for each topic and analyzes

and expands search queries issued by the user: these

tasks are performed through simple and efﬁcient tech-

niques, allowing the system to scale up to a large vol-

ume of user documents.

To obtain an objective, quantitative and repeat-

able evaluation of the goodness of the method, we

employ a systematic assessment process using prede-

ﬁned benchmark data. We employ resources gener-

ally used in IR and freely available on the Web: a col-

lection of Web pages, a search engine indexing them

and a set of test queries with known relevant docu-

ments. For each test query, the relevance of results is

measured with multiple performance indicators. Tests

are run on the raw search engine and on different con-

ﬁgurations of the personalization system, in order to

assess its potential beneﬁts.

The article is organized as follows. Section 2 dis-

cusses and compares against similar methods found

in literature. Section 3 describes the proposed search

personalization method. Section 4 explains our objec-

tive evaluation process, whose outcomes are reported

in Section 5. Finally, Section 6 sums up our work and

suggests possible future directions.

2 RELATED WORK

Personalization of users’ experience in Web search

has been widely explored, addressing both contextu-

alization of search according to the task being car-

ried on and individualization of the experience ac-

cording to the user executing the task (Pitkow et al.,

2002). Earlier approaches used relevance feedback to

reﬁne results upon users’ feedback (Salton and Buck-

ley, 1997), while subsequent are often based on get-

ting implicit feedback by monitoring users’ behavior.

Personalized search (hence PS) generally works

by building a proﬁle of the user and employing it

to customize the search experience. Outride (Pitkow

et al., 2002) deﬁnes a general architecture of a per-

sonalization system mediating between the user and

a general search engine. User-provided query can be

reﬁned to a more precise expression of the informa-

tion need (query rewriting), while results provided by

the engine can be ﬁltered and reordered to promote

those deemed to effectively interest the user. Many

works follow this scheme, differing in how the user

model is built and used (Gauch et al., 2007; Steichen

et al., 2012; Ghorab et al., 2013).

Topics are often a key element in user model-

ing: concept-based proﬁles represent arbitrary mixes

of different topics of interest. Possible topics are

often given by a predeﬁned taxonomy like DMOZ

(discussed in Section 4.1). The OBIWAN project

(Pretschner and Gauch, 1999; Gauch et al., 2003) is

an early example: user’s interests are located on a hi-

erarchy of about 4,400 nodes used as search contexts.

Many works, including ours, combine keywords-

based representation with a concept-based approach:

a user proﬁle is represented by a multiple keyword

vector for each topic of interest. (Trajkova and Gauch,

2004) describe how to proﬁle users by classifying vis-

ited Web pages in a DMOZ-based ontology. In (Liu

et al., 2002) a user proﬁle is used in conjunction with

a global proﬁle; for each query, the most likely top-

ics are proposed to the user. In (Speretta and Gauch,

2005) a list of weighted concepts is extracted from

search results clicked by the user.

Among other existing approaches to user proﬁl-

ing, there are semantic networks, represented by inter-

linked keywords and/or concepts, useful to explicitly

address polysemy and synonymy issues (Micarelli

and Sciarrone, 2004).

The user proﬁle can be built from different infor-

mation sources, like user’s behavior while browsing

the Web and information from local history or cache

(Sugiyama et al., 2004). Some solutions leverage

search history, possibly detecting the most interest-

ing results: for example (Speretta and Gauch, 2005)

gather titles and summaries of results clicked by the

user, while in (Stamou and Ntoulas, 2009) the full

content of selected search results is analyzed.

Few methods explicitly use locally stored user’s

documents, usually merging them with other informa-

tion. In (Teevan et al., 2005), both Web and local doc-

uments are similarly indexed to build models for Web

search personalization: even approximate client-side-

built models can improve relevance of search results.

In (Chirita et al., 2007) is proposed to extract key-

words from the user’s “personal information reposi-

tory”, including text ﬁles as well as e-mails, cached

Web pages and other data. Different techniques to de-

termine the correct additional terms are evaluated and

a method to adaptively choose the number of these

terms is investigated.

These methods aim to build the user’s model upon

large amounts of unlabeled data to build the user’s

model, ignoring the topic of each document. In con-

trast, considering the generally higher value of labeled

data (Domeniconi et al., 2017), our approach focuses

solely on personal ﬁles organized in a hierarchy of di-

rectories: the quantity of analyzable documents may

be smaller, but the use of correctly labeled data can

potentially bring signiﬁcant improvements in the ac-

curacy of the resulting model and consequently in the

relevance of personalized search results.

Other PS solutions based on user-labeled data ex-

ist, like social tagging systems where users can as-

sign free-form tags to Web resources, thus obtaining a

very large corpus of tagged pages (Zhou et al., 2012).

Our solution works with fewer labeled data, but it is

speciﬁcally representative of the target user and its hi-

erarchical nature is useful to distinguish different lev-

els of detail; also, our solution reduces users’ privacy

concerns and removes additional server-side effort.

3 METHOD DESCRIPTION

After formalizing the considered context, we describe

how local knowledge is extracted and used for query

expansion. Figure 1 shows the interconnections be-

tween the components of the method.

3.1 Application Context

Our personalization system works on top of any

search engine Σ responding to any query q with a list

Σ(q) of relevant documents. We formally consider

each query as either a single keyword or a set of sub-

queries combined by AND (∧) or OR (∨) operators.

A collection UD =

F , D, ≺,C

of local user doc-

uments is deﬁned by a set F of ﬁles, a tree D of di-

World

Wide Web

Engine

Local PC

Web

Browser

PS system

Query Expansion

Context Mapping

User

Profile

Text

Processing

Local

User

Files

Aggregation

user query

results

rewritten

query

vector

context keywords

text documents

document

vectors

context

vectors

context

vectors

Figure 1: General diagram of the proposed method. The

upper part of the PS system concerns the construction of the

user proﬁle, while the lower part deals with its application.

rectories partially ordered by the antisymmetric, tran-

sitive relation ≺ ⊂ D × D and a function C : F → D

mapping each ﬁle to the directory containing it. We

denote with L(d) the set of ﬁles directly contained in

the directory d and with L

(d) the set of ﬁles con-

tained in d itself or a subdirectory within.

L(d) = { f ∈ F : C( f ) = d} (1)

(d) = { f ∈ F : C( f ) = d ∨C( f ) ≺ d} (2)

This model can represent a user “home” direc-

tory, containing his or her personal documents conve-

niently arranged in sub-directories. For the purposes

of our method we only consider text ﬁles.

3.2 Extraction of Local Knowledge

We describe how to build a local knowledge base U

constituted by a set of contexts, each representing a

topic. Each context corresponds to a directory d ∈ D.

A context is constituted by the words which represent

the modeled topic, extracted from the text documents

in the corresponding directory.

Each user ﬁle in F is ﬁrst pre-processed to extract

single words contained within it, ﬁltered using stan-

dard IR techniques: casefolding, stopwords removal

and stemming. From each document we obtain a mul-

tiset of ﬁltered words or terms extracted from it. We

denote with #(t, f ) the number of occurrences of term

t in ﬁle f ; for convenience, t ∈ f indicates that t ap-

pears at least once in f , i.e. #(t, f ) > 0.

Now, knowing the set T of distinct terms found

within the ﬁles, we represent each ﬁle f ∈ F with a

vector w

= (w

f ,1

, . . . , w

f ,|T |

) ∈ R

|T |

, indicating the

relevance of each term within f . Such relevance is

computed according to one of the possible weighting

schemes explained shortly.

A vector c

for each directory d ∈ D is trivially

computed as the sum of vectors of every single ﬁle

within the directory itself.

∑

f ∈L(d)

(3)

The set of these vectors constitutes the user pro-

ﬁle U : D → R

|T |

: in fact, each vector corresponds

to a topic in the hierarchy of user interests and indi-

cates which are the most relevant words for that topic.

These vectors will be used for query expansion.

3.3 Term Weighting

Term weighting schemes are employed to properly es-

timate the relevance of terms in single documents.

The use of suitable weighting schemes can be consid-

erably beneﬁcial in many practical contexts (Domeni-

coni et al., 2016). The weight of a term within a doc-

ument f is generally composed of a local factor based

solely on the contents of f and a global factor com-

puted on the whole set of documents.

As the local weight factor for a term t in f , known

as term frequency (tf ), we use the number of occur-

rences of t normalized according to the maximum

number of occurrences of a term in f itself.

tf(t, f ) =

#(t, f )

max

τ∈ f

#(τ, f )

(4)

The most common choice for global factor would

be the inverse document frequency (idf), assigning

higher weights to terms appearing in fewer docu-

ments. However, taking inspiration from supervised

term weighting schemes used in automated text cate-

gorization to compute different weights for each topic

category (Domeniconi et al., 2016), we also consider

the directory where a ﬁle is located when computing

weights for its terms.

A ﬁrst supervised global measure we propose,

called idfd (idf in directory), is in practice the stan-

dard idf measured only on documents of the direc-

tory being considered, recursively including its sub-

directories. This allows to evaluate the discriminative

power of a term within the speciﬁc topic being con-

sidered, rather than in the whole taxonomy.

idfd(t, d) = log



(d)|

| f ∈ L

(d) : t ∈ f |



(5)

The second option we propose, called idfod (idf

outside of directory), is to instead compute idf ex-

cluding documents in the considered directory and its

nested subdirectories. The rationale is to avoid under-

estimating terms which appear frequently in the con-

sidered directory, but rarely elsewhere.

idfod(t, d) = log



|F \ L

(d)|

| f ∈ F \ L

(d) : t ∈ f |



(6)

Either one G of these global factors is multiplied

by the local factor (tf) to obtain the ﬁnal weight of a

term t in a ﬁle f in a directory d.

tf · G(t, f , d) = tf(t, f ) · G(t, d) (7)

3.4 Query Expansion

To expand a query, we have ﬁrst to determine its effec-

tive context among the ones inferred from the user’s

directories. One option would be to let the user man-

ually select a context, as in e.g. (Liu et al., 2002): this

requires some more effort from the user, although he

or she may prefer to pick a category from a limited set

rather than typing additional search keywords.

However, to reduce the user’s effort, we use a sim-

ple method to automatically map each query to its

most likely context. Using the same pre-processing

steps and term weighting schemes employed for user

documents, a query q is converted to a vector q ∈ R

|T |

in the same form of the contexts. In this way, employ-

ing the common cosine similarity measure between

vectors, we pick the context most similar to the query.

C(q) = argmax

d∈D

q · c

||q|| · ||c

(8)

Once the context of the query is determined, the

actual expansion can be performed. At an high level,

this involves determining a set K = {k

, k

, . . .} of

keywords and adding them to the user’s query.

As the set of keywords, the n

terms with the high-

est weight in the vector of the picked context is used.

> 0 is a parameter of the method.

Finally, for any query q given by the user, the

expanded query q

is obtained from the conjunction

(AND) between q and the disjunction (OR) of all the

keywords picked from the context. This new query q

is sent to the search engine in place of q.

= q ∧

k∈K

k = q ∧ (k

∨ k

∨ . . .) (9)

4 EVALUATION PROCESS

To objectively assess our solution for personalized

search, we set up a repeatable evaluation process us-

ing available benchmark data, providing quantitative

measures of the relevance of results given by a search

engine. We can thus measure how much our method

can improve the relevance of topmost search results

when applied to a “plain” search engine.

While methods proposed for personalized search

are usually accompanied by the results of an exper-

imental evaluation of their goodness, comparisons

across different methods are generally unfeasible, due

to the use of different evaluation processes, particu-

larly when they are not objective, or because of data

sets not made available by the authors. Especially,

many works (e.g. (Pitkow et al., 2002; Micarelli and

Sciarrone, 2004)) employ a user-oriented assessment,

where the system is tested by an heterogeneous group

of people: a feedback about usability is obtained

through questionnaires or by tracking time and ac-

tions required by users for each query. This approach

has the advantage of estimating how much the tested

system actually improves the search experience, es-

pecially in terms of personalization to the different

users. However, for the sake of objectiveness and

ease of repeatability, we favor the quantitative mea-

surement of retrieval effectiveness as described in the

following.

For the systematic evaluation of an IR system, we

ﬁrst need a corpus P of documents to be indexed.

To evaluate the system behavior, we then use a set

Q = {q

, q

, . . .} of queries: in each test, the sys-

tem responds to a query with a relevance-sorted list

of documents picked from P . To evaluate their cor-

rectness, we need a gold standard to compare against:

this is constituted, for each query q, by an expert-

made relevance judgment on each document in P . For

each query, a perfect IR system should return all the

relevant documents and nothing else. As users often

look only at the topmost results of a search engine,

we only consider the ﬁrst 20 results for each query

response: we use these results and the gold standard

given for each query to compute four different perfor-

mance measures, as explained later.

Since personalized search works around an exist-

ing IR system, another necessary element is a refer-

ence search engine upon which to apply our method.

Similarly to other works, we perform an evaluation

on the plain search engine to obtain baseline results,

which are then compared to those obtained by apply-

ing personalization on top of the same engine.

To test a personalized search system, other than

the “global” test data described above, is also needed

either an example user proﬁle or data useful to build

it. In our case, to build a proﬁle, we employ a corpus

of text documents organized in a hierarchy of directo-

ries, reﬂecting the personal ﬁles of a generic user.

4.1 Benchmark Data

The benchmark corpus of Web pages we use is

extracted from the ClueWeb09 Dataset by Lemur

Project

: our collection is constituted by the

503,903,810 pages in English language.

http://www.lemurproject.org/

Table 1: Sample queries used in experimental evaluation.

horse hooves uss yorktown charleston sc

avp ct jobs

discovery channel store penguins

president of the united states how to build a fence

iron bellevue

To simulate a Web search engine indexing these

pages we use the Indri search engine, also by Lemur

Project, providing a ﬂexible query language including

the AND and OR operators used in query expansion.

The IR system, either personalized or not, has to

be tested with a set of sample search queries rep-

resenting different information needs and having a

known set of actually relevant documents within the

indexed ones. We employ data used in the 2010

Web Track

of Text REtrieval Conference (TREC)

for evaluation of participant IR systems, based on the

ClueWeb09 collection. For the competition, NIST

created 50 search topics, each including a query string

likely to be searched by a user interested in that topic:

10 of these queries are listed in Table 1. To each topic

is associated a set of documents within ClueWeb09

judged as relevant to it. Each relevant document is

also graded in an integer scale ranging from 1 (little

information about the topic) to 4 (very speciﬁc infor-

mation).

Finally we need a corpus of text documents resem-

bling ﬁles stored on a user’s PC: documents should

discuss different topics and be accordingly organized

in a hierarchy of directories. We take advantage of

DMOZ (http://dmoz.org), an open-content directory

of about four million Web sites organized in a tree-

like hierarchy of about one million categories, rang-

ing from generic, e.g. “Science”, to speciﬁc ones such

as “Bayesian Analysis”. We extract a data set consti-

tuted by the ﬁrst two levels of the hierarchy, excluding

the “World” and “Regional” top-level categories con-

taining non-English pages. The resulting set of 5,184

Web pages is considered as a collection of personal

documents, with the 13 top-level and 308 nested cat-

egories treated as directories.

We choose to only consider the two topmost lev-

els of DMOZ basing on the hypothesis that the aver-

age user employs a hierarchy of the same depth, with

good balance between speciﬁcity of topics and ease

of organization. A deeper taxonomy would require

more user effort and would be generally worthless for

the typical amount of documents stored locally.

While the spectrum of topics considered in this

data set is considerably wide compared to the set of

interests of a typical user, we claim that it is suitable

in our evaluation process. In fact, as the test queries

http://trec.nist.gov/data/web10.html

cover a wide range of different topics, the corpus of

personal documents should include them all in order

to have the chance of picking a suitable context. In a

real use case, the local user’s ﬁles would likely in-

clude a narrower set of topics, but also the issued

queries would still mostly ﬁt into the same set: our

selection of benchmark data reﬂects this latter point.

4.2 Performance Measures

We consider four different performance measures

proposed in literature, used to evaluate IR systems in

the TREC 2010 Web Track cited above. Each mea-

sure is computed for each query q by comparing the

list Σ(q) = (r

, r

, . . .) of search results to known rel-

evance judgments; results for each measure are aver-

aged across test queries. We denote with Σ

(q) the list

Σ(q) truncated at the topmost k entries, where k = 20

in our evaluation.

The ﬁrst two measures are simply based on a bi-

nary labeling of test documents as either “relevant” or

“not relevant” to each query. R

⊂ P denotes the set

of documents relevant to the query q.

When search results are in the form on unranked

sets, the standard precision and recall measures are

usually employed: especially, the precision is the

ratio of actually relevant documents among the re-

trieved ones. In the case of a ranked list of results,

it is common to measure the precision only on the

topmost k results (with k usually within 10 and 30),

termed precision@k or P@k for short.

P@k(q) =

|Σ

(q) ∩ R

|Σ

(q)|

(10)

The Mean Average Precision (MAP) is another

commonly used measure for evaluation in IR. The av-

erage precision (AP) on a query is obtained by averag-

ing for each relevant document the precision@k with

k equal to its position in the results, considering 0 for

documents excluded from the results. MAP is simply

the average of AP on all test queries.

AP(q) =

∑

∈R

P@i(q) (11)

The following two measures support a non-binary

notion of relevance: rather than marking a document

r as either “relevant” or “not relevant” to a query q,

a real relevance score R(q, r) ∈ [0, 1] evaluates how

much r satisﬁes the information need expressed by q.

In our case, where relevance of r to q is evaluated

by a grade g ∈ {0, 1, 2, 3, 4} with 0 indicating a non-

relevant document, the score is computed as follows.

R(q, r) =

− 1

(12)

One measure using such relevance scores is the

normalized discounted cumulative gain (NDCG),

computed in the formula below, where Z is a normal-

ization factor ensuring that the maximum achievable

NDCG@k is 1.

NDCG@k(q) = Z

∑

m=1

R(q,r)

− 1

log

(1 + m)

(13)

Finally, the expected reciprocal rank (ERR)

(Chapelle et al., 2009), contrarily to other measures, is

based on the so-called cascade model where the user

just searches for a single relevant document, starting

from the top of the results. The ERR measures the

expected number of documents the user has to check

before ﬁnding the ﬁrst relevant one.

ERR@k(q) = Z

∑

m=1

R(q, r

)

m−1

∏

n=1

(1 − R(q, r

)) (14)

5 EXPERIMENTAL RESULTS

We compare different test conﬁgurations, testing in

each different combinations of values for the consid-

ered parameters, which are the used term weighting

scheme and the number n

of maximum keywords

used to expand each query. Table 2 summarizes the

results, grouped by conﬁguration. For every conﬁg-

uration and for both tf-idfd and tf-idfod weighting

schemes, we report the best results obtained by vary-

ing the n

parameter between 1 and 50, along with

the lowest value of the parameter which yielded the

reported measure.

In the ﬁrst test no personalization is applied: test

queries are issued as they are to the Indri search en-

gine and the considered performance metrics are mea-

sured by comparing the topmost 20 results for each

query with the known relevant documents. These re-

sults serve as a baseline for tests on the personaliza-

tion system: for each measure, both its absolute value

and the relative gain on the baseline are considered.

To test our approach, we ﬁrst devise an ideal case

where the local user documents perfectly match the

information needs expressed by the test queries. In

practice, we suppose that the user’s home directory

has a subdirectory for each test query q ∈ Q exactly

containing the relevant documents R

for it. We also

assume that the system maps each query to the corre-

sponding directory. While this conﬁguration is obvi-

ously unrealistic, these tests show the plausible maxi-

mum potential of our personalization approach.

From the results obtained with this conﬁguration

(section A of Table 2), we see that our personalization

system, in ideal conditions, could signiﬁcantly boost

the relevance of search results.

Table 2: Relevance measures for different search conﬁgurations with the best values of the n

parameter.

conﬁguration weighting MAP P@20 ERR@20 nDCG@20

scheme n

acc. impr. n

acc. impr.

baseline - - 0.073 - - 0.180 - - 0.062 - - 0.073 -

A tf-idfd 18 0.137 87.7% 8 0.360 100.0% 11 0.172 177.4% 8 0.195 167.1%

perfect mapping tf-idfod 20 0.127 74.0% 20 0.270 50.0% 14 0.116 87.1% 14 0.130 78.1%

B tf-idfd 20 0.094 28.8% 16 0.275 52.8% 15 0.104 67.7% 16 0.120 50.7%

optimal mapping tf-idfod 24 0.064 -12.3% 24 0.185 2.8% 24 0.081 30.6% 24 0.077 5.5%

C tf-idfd 23 0.078 6.8% 9 0.220 22.2% 18 0.095 53.2% 18 0.084 15.1%

automatic mapping tf-idfod 15 0.090 23.3% 17 0.230 27.8% 15 0.090 45.2% 17 0.084 15.1%

In the next two more realistic conﬁgurations, doc-

uments extracted from DMOZ as described in Section

4.1 are used as the corpus of local user documents, so

that there is no match between the documents indexed

by the search engine and those upon which the user

proﬁle is built. Also, test queries do not directly cor-

respond to user directories anymore, thus requiring a

non-trivial mapping between them.

We ﬁrst devise a case where this mapping is super-

vised: the correct context of each test query is the one

under which most actual relevant documents are clas-

siﬁed by a k-NN classiﬁer trained on the ODP docu-

ments used in the local ﬁles collection. Despite this

process would not be feasible in a real case where

the relevant documents are not known in advance,

it provides an indication of how much a proper re-

formulation of a query would improve search results

and simulates cases where the query-directory map-

ping is based on knowledge beyond the query itself.

This would happen for example if the user explicitly

indicates the intended context of the query, picking

among his or her directories.

Results for this conﬁguration (section B of Table

2), while notably inferior to the previous ideal ones,

still show signiﬁcant improvements over the baseline,

especially where the tf-idfd term weighting scheme

is employed. On the other side, use of the tf-idfod

scheme does not notably improve results in this case,

which sometimes are even worse than the baseline.

Lastly, we consider the actual procedure described

in Section 3.4: for each query, the context most sim-

ilar to the query itself is selected, without employing

additional knowledge. This is intended to be the most

common use case, where the user’s input is limited

to a short and possibly ambiguous query and the per-

sonalization system ought to automatically infer the

correct context to be used to expand it.

Results on this conﬁguration (section C of Table

2) show an overturning regarding the term weight-

ing schemes. While tf-idfd brought better results

than tf-idfod in previous conﬁgurations, here gener-

ally brings scarce improvements over the baseline.

On the contrary, tf-idfod performs best according to

most performance measures, even improving the re-

sults obtained with the previous conﬁguration, where

optimal query-directory mapping was used. In any

case, more or less signiﬁcant improvements over the

no-personalization baseline are obtained, especially

using ERR@20 as the reference measure.

Regarding the number n

of keywords added to

each query to personalize it, in most cases its opti-

mal value ranges between 15 and 20. Numbers in this

interval seem thus to constitute the correct balance be-

tween providing enough terms to effectively make the

query unambiguous and avoiding to add terms which

are not relevant enough to the context.

6 CONCLUSIONS

We proposed a method to personalize Web search by

expanding outgoing queries according to the speciﬁc

user’s interests. With respect to known methods, the

major novelty of our solution consists into leverag-

ing text ﬁles locally stored by the user and speciﬁ-

cally considering their natural hierarchical organiza-

tion into directories. From this collection of docu-

ments already subdivided by topics, we can extract

representative keywords for each topic, infer the topic

of each query issued by the user and expand it by

adding keywords indicating the topic.

To obtain an objective assessment of the effective-

ness of our solution, we set up a systematic evaluation

process based on predeﬁned benchmark data. A refer-

ence search engine has been tested on a ﬁxed corpus

of Web pages using a set of test queries, in order to ob-

tain baseline indicators of the retrieval performance.

The same evaluation has then been applied on dif-

ferent conﬁgurations of our personalization method,

running upon the same reference search engine.

Our experiments show that query expansion could

boost signiﬁcantly the relevance of topmost results

returned by a search engine and that our method is

able to improve relevance with information extracted

solely from the user’s local documents. While the

user could manually select the correct context of each

query in order to maximize accuracy, the trivial au-

tomatic mapping still guarantees interesting improve-

ments over the use of the plain search engine.

The method could be reﬁned in many ways. In this

work we used simple and well-established procedures

to extract knowledge from hierarchically-organized

local documents and to expand search queries: the ex-

perimentation of more advanced techniques may lead

to improvements of the proposed general method.

While we focused on enhancing user’s queries, the

personalization system could also work on the out-

put of the underlying search engine, ﬁltering or re-

ordering results better matching known user’s inter-

ests: this approach could be used in substitution of or

even in combination with query expansion, exploiting

the same user model. Finally, methods could be de-

vised to make use of local user’s documents alongside

other information sources, such as search or browsing

history.

REFERENCES

Carpineto, C. and Romano, G. (2012). A survey of auto-

matic query expansion in information retrieval. ACM

Computing Surveys (CSUR), 44(1):1.

Chapelle, O., Metlzer, D., Zhang, Y., and Grinspan, P.

(2009). Expected reciprocal rank for graded rele-

vance. In Proceedings of the 18th ACM conference on

Information and knowledge management, pages 621–

630. ACM.

Chirita, P.-A., Firan, C. S., and Nejdl, W. (2007). Person-

alized query expansion for the web. In Proceedings

of the 30th annual international ACM SIGIR confer-

ence on Research and development in information re-

trieval, pages 7–14. ACM.

Domeniconi, G., Moro, G., Pagliarani, A., and Pasolini, R.

(2017). On deep learning in cross-domain sentiment

classiﬁcation. In Proceedings of the 9th International

Joint Conference on Knowledge Discovery, Knowl-

edge Engineering and Knowledge Management.

Domeniconi, G., Moro, G., Pasolini, R., and Sartori, C.

(2016). A comparison of term weighting schemes for

text classiﬁcation and sentiment analysis with a super-

vised variant of tf.idf. In Data Management Technolo-

gies and Applications (DATA 2015), Revised Selected

Papers, volume 553, pages 39–58. Springer.

Gauch, S., Chaffee, J., and Pretschner, A. (2003). Ontology-

based personalized search and browsing. Web Intelli-

gence and Agent Systems: An international Journal,

1(3, 4):219–234.

Gauch, S., Speretta, M., Chandramouli, A., and Micarelli,

A. (2007). User proﬁles for personalized information

access. In The adaptive web, pages 54–89. Springer.

Ghorab, M. R., Zhou, D., OConnor, A., and Wade, V.

(2013). Personalised information retrieval: survey and

classiﬁcation. User Modeling and User-Adapted In-

teraction, 23(4):381–443.

Jansen, B. J., Spink, A., and Saracevic, T. (2000). Real

life, real users, and real needs: a study and analysis

of user queries on the web. Information processing &

management, 36(2):207–227.

Liu, F., Yu, C., and Meng, W. (2002). Personalized web

search by mapping user queries to categories. In Pro-

ceedings of the eleventh international conference on

Information and knowledge management, pages 558–

565. ACM.

Micarelli, A. and Sciarrone, F. (2004). Anatomy and empir-

ical evaluation of an adaptive web-based information

ﬁltering system. User Modeling and User-Adapted

Interaction, 14(2-3):159–200.

Pitkow, J., Sch

utze, H., Cass, T., Cooley, R., Turnbull, D.,

Edmonds, A., Adar, E., and Breuel, T. (2002). Person-

alized search. Communications of the ACM, 45(9):50–

55.

Pretschner, A. and Gauch, S. (1999). Ontology based per-

sonalized search. In Tools with Artiﬁcial Intelligence,

1999. Proceedings. 11th IEEE International Confer-

ence on, pages 391–398. IEEE.

Salton, G. and Buckley, C. (1997). Improving retrieval per-

formance by relevance feedback. Readings in infor-

mation retrieval, 24(5):355–363.

Speretta, M. and Gauch, S. (2005). Personalized search

based on user search histories. In Web Intelligence,

2005. Proceedings. The 2005 IEEE/WIC/ACM Inter-

national Conference on, pages 622–628. IEEE.

Stamou, S. and Ntoulas, A. (2009). Search personalization

through query and page topical analysis. User Model-

ing and User-Adapted Interaction, 19(1-2):5–33.

Steichen, B., Ashman, H., and Wade, V. (2012). A compar-

ative survey of personalised information retrieval and

adaptive hypermedia techniques. Information Pro-

cessing & Management, 48(4):698–724.

Sugiyama, K., Hatano, K., and Yoshikawa, M. (2004).

Adaptive web search based on user proﬁle constructed

without any effort from users. In Proceedings of the

13th international conference on World Wide Web,

pages 675–684. ACM.

Teevan, J., Dumais, S. T., and Horvitz, E. (2005). Per-

sonalizing search via automated analysis of interests

and activities. In Proceedings of the 28th annual in-

ternational ACM SIGIR conference on Research and

development in information retrieval, pages 449–456.

ACM.

Trajkova, J. and Gauch, S. (2004). Improving ontology-

based user proﬁles. In Proceedings of RIAO 2004,

pages 380–390.

Zhou, D., Lawless, S., and Wade, V. (2012). Improving

search via personalized query expansion using social

media. Information retrieval, 15(3-4):218–242.