SEARCHING KEYWORD-LACKING FILES BASED ON LATENT

INTERFILE RELATIONSHIPS

Tetsutaro Watanabe

, Takashi Kobayashi

and Haruo Yokota

Department of Computer Science, Graduate School of Information Science and Engineering

Tokyo Institute of Technology, Tokyo, Japan

Department of Information Engineering, Graduate School of Information Science

Nagoya University, Nagoya, Japan

Keywords:

Desktop search, File retrieval, Access log analysis, Latent inter-ﬁle relationship.

Abstract:

Current information technologies require ﬁle systems to contain so many ﬁles that searching for desired ﬁles

is a major problem. To address this problem, desktop search tools using full-text search techniques have been

developed. However, those ﬁles lacking any given keywords, such as picture ﬁles and the source data of

experiments, cannot be found by tools based on full-text searches, even if they are related to the keywords.

It is even harder to ﬁnd ﬁles located in different directories from the ﬁles that include the keywords. In

this paper, we propose a method for searching for ﬁles that lack keywords but do have an association with

them. The proposed method derives relationship information from ﬁle access logs in the ﬁle server, based on

the concept that those ﬁles opened by a user in a particular time period are related. We have implemented

the proposed method, and evaluated its effectiveness by experiment. The evaluation results indicate that the

proposed method is capable of searching keyword-lacking ﬁles and has superior precision and recall compared

with full-text and directory-search methods.

1 INTRODUCTION

Advances in information technologies have led to

many types of multimedia data, including ﬁgures, im-

ages, sounds, and videos, being stored as ﬁles in com-

puter systems alongside conventional textual mate-

rial. Moreover, the recent price drop for magnetic

disk drives (Hayes, 2002) has accelerated the explo-

sive increase in the number of ﬁles within typical ﬁle

systems (Agrawal et al., 2007).

Most current operating systems adopt hierarchical

directories to manage ﬁles. A very large number of

ﬁles make the structure of such a directory very exten-

sive and complex. Therefore, it is very hard to clas-

sify the many ﬁles into appropriate directories. Even

if all the ﬁles are classiﬁed appropriately, it is still dif-

ﬁcult to ﬁnd a desired ﬁle located at a deep node in

the directory tree. The results obtained by Dumais et

al. (Dumais et al., 2003)also lead to that conclusion as

they observe that users feel less the need to maintain a

complex hierarchy of their documents when they have

a powerful search tool allowing them to ﬁnd their doc-

uments more easily.

In this context, several desktop search tools using

full-text search techniques have been developed, such

as Google Desktop, Windows Desktop Search, Spot-

light, Namazu (Namazu, 2009), Hyper Estraier (Es-

traier, 2007). Moreover several ranking techniques

for desktop search systems also have been pro-

posed (Cohen et al., 2008). However, their main tar-

get is restricted to text-based ﬁles such as Microsoft

Word documents, PDFs, and emails.

Other types of ﬁles, such as picture ﬁles, image

ﬁles, and source data ﬁles for experiments and ﬁeld

work, cannot be found by these full-text search tools

because they lack search keywords. Even for text-

based ﬁles, they cannot be found if they do not include

directly related keywords. It becomes even harder if

these ﬁles are located in different directories from the

ﬁles that contain the keywords.

To address the demand for searching for these

keyword-lacking ﬁles, there has been much research

aiming to append metadata to ﬁles (Yee et al., 2003).

However, it is practically impossible to assign “per-

fect” metadata to a large number of ﬁles. On the other

hand, Google Image Search (Google, 2010) is capa-

ble of searching image ﬁles in web sites associated

with the keywords by using the reference information

236

Watanabe T., Kobayashi T. and Yokota H. (2010).

SEARCHING KEYWORD-LACKING FILES BASED ON LATENT INTERFILE RELATIONSHIPS.

In Proceedings of the 5th International Conference on Software and Data Technologies, pages 236-244

DOI: 10.5220/0002931002360244

 SciTePress

in HTML ﬁles within the site. However, this method

does not directly apply to ﬁles in the ﬁle system be-

cause they rarely contain information relevant to the

sources of the included objects.

In this paper, to provide the function of searching

for keyword-lacking ﬁles that match with given key-

words, we focus on the latent relationship between

ﬁles that have been frequently accessed at about the

same time. For example, when a user of a com-

puter system edits the ﬁle for a research paper con-

taining conceptual ﬁgures and graphs of experiments,

the user frequently opens ﬁles for the ﬁgures and

data sources of the experiments at the same time.

Of course, other ﬁles that do not directly relate to

the paper, such as reports for lectures and emails to

friends, are also opened simultaneously. However, the

frequency of opening related ﬁles at the same time

should be higher than the frequency for non-related

ﬁles.

To achieve this function, we propose a method

for mining latent interﬁle relationships from the ﬁle

access logs in a ﬁle server(Watanabe et al., 2008)

. We have implemented a desktop search system

“FRIDAL” based on the proposed method using ac-

cess logs for Samba. FRIDAL is an acronym for

“File Retrieval by Interﬁle relationships Derived from

Access Logs”. To evaluate the method, we com-

pared the search results for FRIDAL with a full-

text search method, directory search methods, and

a related method used in Connections (Soules and

Ganger:, 2005). The evaluation results, using actual

ﬁle access logs for testers indicated that the proposed

method is capable of searching the keyword-lacking

ﬁles that cannot be found by other methods, and that

it has superior precision and recall compared with the

other approaches.

The remainder of this paper is organized as fol-

lows. First, we review related work in Section 2.

Then we propose a method for mining the ﬁle access

logs in Section 3, and describe the implementation of

FRIDAL in Section 4. In Section 5, we compare the

search results for FRIDAL with the other methods to

evaluate FRIDAL. We conclude the paper and discuss

future work in Section 6.

2 RELATED WORK

There are interesting discussions related desktop

searching task. Barreau and Nardi (Barreau and

Nardi:, 1995) summarize and synthesize investigated

information organization practices among users of

several DOS, Windows, OS/2 and Macintosh. Fertig

et al. (Fertig et al., 1996) have refuted Barreau’s ar-

guments on the desktop and ﬁle & folder metaphor,

which is analogy to our paper-based world. They

mentioned several approaches included the virtual di-

rectories of the MIT Semantic File System (Gifford

et al., 1991), Lifestreams (Freeman and Gelernter,

1996), which uses a time-based metaphor and fast

logical searches to organize, monitor, ﬁnd and sum-

marize information. Blanc-Brude and Scapin (Blanc-

Brude and Scapin, 2007) also has discussed which at-

tributes people actually recall about their own docu-

ments, and what are the characteristics of their recall.

The MIT Semantic File System (Gifford et al., 1991)

enables a ﬁle to have a number of attributes, instead of

placing it in a directory, to help share its information.

The method proposed in the article (Ishikawa et al.,

2006) uses the Resource Description Framework to

describe the attributes of ﬁles, and applies seman-

tic web technology to manage ﬁles by maintaining

their consistency. Chirita et al. (Chirita et al., 2005)

were also proposed semantic approaches for desktop

search. Our method and these other approaches have

the same goal; solving the problem of the hierarchi-

cal directory. However, our approach uses ﬁle access

logs instead of semantics.

There also exists some research which uses a time-

based metaphor. In time-machine computing (Reki-

moto, 1999), all ﬁles are put on the desktop and grad-

ually disappear over time. If a user inputs a date

to the system, the user can see the desktop at the

appointed date. The system also supports keyword

searches by using an “electronic post-it” note created

by the user. Dumais et al. have proposed “Stuff

I’ve Seen”(SIS) (Dumais et al., 2003) which supports

to ﬁnding and re-using previously seen information

on desktop. In SIS, information that a person has

seen is indexed and he can search by rich contex-

tual cues such as time, author, thumbnails and pre-

views. Their time-centric approach differs from our

keyword-centric approach.

OreDesk (Ohsawa et al., 2006) derives user-active

records from OS event logs and installed plug-ins.

Then it calculates user-focused degrees for web pages

and ﬁles (called Datas) based on the active records,

and also calculates relationships between Datas. It

provides a search function for related ﬁles, given a

Data name, and also provides a viewer for Datas and

relationships. Whereas OreDesk changes the user’s

environment to derive the active records, our method

does not change it because we use the access logs

for the ﬁle server. Also, whereas OreDesk uses the

Data name for searching, our method uses keywords.

Furthermore, OreDesk takes account of the start time

of Data only, using it to calculate the relationships,

whereas our method takes account of total time, num-

SEARCHING KEYWORD-LACKING FILES BASED ON LATENT INTERFILE RELATIONSHIPS

237

ber, and separation of co-occurrences.

Chirita and Nejdl have discussed authority trans-

fer annotation (Nejd and Paiu, 2005) and proposed

a ranking method for desktop search using PageR-

ank (Page et al., 1999) on connections by exploit-

ing usage analysis information about sequences of ac-

cesses to local resources (Chirita and Nejdl, 2006).

They have also proposed a method for use ﬁle us-

age analysis as an input source for clustering desktop

documents (Chirita et al., 2006). A distance between

documents have been deﬁned by using the number of

steps between consecutive accesses of ﬁles and a time

window in which they occur.

Connections (Soules and Ganger:, 2005) obtains

system calls to ﬁles, such as read() and write(), and

constructs a directed graph comprising the ﬁles (as its

nodes) and the relationships (as its edges). Following

a search request, it performs a context-based search,

and then searches for related ﬁles in the result of the

context-based search, by tracing the directed graph.

The aim of Connections is the same as that of our

method. However, whereas Connections aims to de-

rive the reference/referenced relationships via system

calls, our method derives information of ﬁle usage via

open-ﬁle/close-ﬁle information in access logs. It also

differs in calculating the relationships of ﬁles in the

search results. Since the logs of Samba are obtained

without any modiﬁcations for the target system, our

method is easier to be implemented than the method

based on system-calls’ logs. Moreover, overhead of

obtaining logs in a ﬁle server is very small compared

with the system calls.

We will describe the calculations of Connections

because we will be comparing our method to the cal-

culations of Connections in Section 5. In (Soules and

Ganger:, 2005), various calculation methods are pro-

posed. We will explain the most efﬁcient of these, as

reported by the authors. First, Connections makes an

edge whose weight is 1 from the read ﬁle to the writ-

ten ﬁle in a time window speciﬁed by the constant

TimeWindows. If the edge already exists, its weight

is incremented by 1. In the search phase, it ﬁrst per-

forms a context-based search. Let w

be the point of

the ﬁles, as scored by the context-based search. If the

ﬁle is not to be included in the result, then w

= 0.

Let E

be the set of edges going to the node m, where

∈ E

is the weight of the edge from n to m divided

by the sum of the weights of the edges leaving node

n. If e

<Weight Cutoff, then the edge is ignored.

The point of node w

is calculated as follows when

the number of repeats is P.

∑

∈E

(i−1)

· [α· e

+ (1− α)] (1)

∑

i=0

(2)

3 PROPOSED METHOD

In this paper, we propose a method capable of search-

ing keyword-lacking ﬁles using latent interﬁle rela-

tionships, which has both a preparing and a searching

phase.

Preparing Phase. Derive information about ﬁles us-

age from ﬁle access logs, and calculate latent in-

terﬁle relationships.

Searching Phase. Obtain keywords, perform a full-

text search using the keywords, calculate the point

of ﬁles related to the results of the full-text search

based on the latent interﬁle relationships, and

show the ﬁles ordered by point (Figure 1).

The proposed method assumes that the ﬁles are

stored in the ﬁle server, and are used on personaldesk-

tops. In this paper, we assume that the ﬁles are not

copied, moved, or renamed.

Figure 1: Overview of the proposed method.

3.1 Preparing Phase

We focus on the relationship between ﬁles that are fre-

quently used at the same time. When a user edits a ﬁle

during a task, the user often opens related ﬁles to re-

fer to or to edit. In this case, the same ﬁles are used

frequently every time the task is performed. There-

fore, we suppose that ﬁles that are frequently used in

similar timing have a latent relationship.

In our research, to identify such ﬁles, we derive

information about ﬁle usage from ﬁle access logs.

3.1.1 Derive Approximate File Use Duration

First, we obtain the “open-ﬁle” and “close-ﬁle” for

all ﬁles for all users from the ﬁle access logs for a

ﬁle server. Next, we deﬁne the File Use Duration

“FUD” as the time between open-ﬁle and close-ﬁle.

ICSOFT 2010 - 5th International Conference on Software and Data Technologies

238

If a close-ﬁle is not recorded, we remove the corre-

sponding open-ﬁle. For access logs in Samba, there

are two problems that mean that the actual duration of

ﬁle use differs from the FUD. We propose a method

for solving these problems to obtain an Approximate

File Use Duration “AFUD”, as shown in Fig2). Fi-

nally, we ﬁlter out any AFUD that has a time span

less than T

Figure 2: Derive approximate ﬁle use duration.

We now describe the two access log problems and

how to solve them.

Problem A. A user sometimes leaves his or her seat

with ﬁles open.

Problem B. There are two types of applications. One

application locks a ﬁle when it opens ﬁle, whereas

the other copies it to memory and does not lock

it. The latter type causes a difference between the

FUD and the actual duration.

To solve these problems, we prepare the follow-

ing two pieces of information by analyzing the access

logs.

Active Time List. This list comprises information

about the existence of logs in all time windows

. “Logs exist” means a user was active at that

time.

File Type List. This is a list of ﬁle types whose av-

erage FUD is less than T

. An item in this list is

possibly used by an application that unlocks ﬁles.

For Problem A, we mask the FUD with the Active

Time List. Then, if the span of active time is greater

than T

, we delete the corresponding FUD, as shown

in Figure 3, because we think the user forgot to close

the ﬁle and went home. For Problem B, for the ﬁle

types in the File Type List, we extend the FUD from

the ﬁrst FUD to the last FUD for each item in the

Active Time List, as shown in Figure 4, because we

think the ﬁrst FUD is “open” and the last FUD is “save

and close”.

3.1.2 Calculate Latent Interﬁle Relationships

We assume that strongly related ﬁles are used at the

same time when executing the same task. We calcu-

late the latent interﬁle relationships by the AFUDs de-

Figure 3: Managing Problem A.

Figure 4: Managing Problem B.

scribed in the previous section. To achieve this, we in-

troduce four “relationship elements”, where the term

“CO” (meaning co-occurrence) is deﬁned as the over-

lap of two AFUDs.

T : Total time of COs.

C : Number of COs.

D : Total time of the time span between COs.

P : Similarity of the timings of the open-ﬁle opera-

tions.

By using these four relationship elements and four

parameters, α, β, γ, and δ, we deﬁne the latent interﬁle

relationship R as follows.

R = T

·C

· D

· P

(3)

Figure 5: The values used in the calculation of relationship

elements.

SEARCHING KEYWORD-LACKING FILES BASED ON LATENT INTERFILE RELATIONSHIPS

239

We now describe how to calculate relationship ele-

ments, by reference to Figure 5. Let {z

, z

, ··· , z

}

be n COs of two ﬁles, let {t

, · ·· , t

} be their times,

let {d

, d

, · ·· , d

(n−1)n

} be the time spans between

each CO, and let {p

, p

, · ·· , p

} be the difference in

the start time of two AFUDs. Each relationship ele-

ment is calculated as follows.

T =

∑

i=1

D =

(

1 n = 0

∑

n−1

i=1

i(i+1)

otherwise

C = n P =

(

1 ∀i p

= 0



∑

i=1



−1

otherwise

3.2 Searching Phase

In the searching phase, we ﬁrst run the full-text search

using the input keywords. Then we score the ﬁle point

by using the TF–IDF (Term Frequency–Inverse Doc-

ument Frequency) and latent interﬁle relationships for

all ﬁles related to the ﬁles found in the full-text search.

Figure 6: Calculation of the point of ﬁles.

The ﬁle point is calculated as a high value when

the ﬁle has a strong relationship with a high TF–IDF

ﬁle. Let U be the set of all ﬁles, and let R( f

, f

) be

the latent interﬁle relationship between ﬁle f

and ﬁle

For the case of a single input keyword, we run a

full-text search using the keyword k. Let F(k) be the

resulting set of ﬁles. Then, for each ﬁle f ∈ F(k), we

calculate the TF–IDF S

( f, k) of f using k. Note that

the TF–IDF is zero if a ﬁle is not in the full-text search

result. Next, we normalize R( f

, f

) to

R( f

, f

, k) un-

der keyword k as follows.

R( f

, f

, k) =

log[R( f

, f

)]

log[ max

f∈F(k)

′

∈U

{R( f, f

′

)}]

(4)

In this normalization, the highest possible rela-

tionship degree becomes 1. Then the added point

( f

, f

, k) to ﬁle f

, based on the relationship with

, is calculated as follows.

( f

, f

, k) =

R( f

, f

, k) ·S

( f

, k) (5)

Finally, the point S( f

, k) of ﬁle f

is calculated as

follows.

S( f

, k) = S

( f

, k) +

∑

f∈F(k)

( f

, f, k) (6)

We can explain this by using a concrete exam-

ple. Refer to the left-hand side of Figure 6. Af-

ter searching using the keyword k, f

, f

are

found to include the keyword. The TF–IDFs for these

ﬁles are S

( f

, k) = 10, S

( f

, k) = 20, S

( f

, k) =

0, and S

( f

, k) = 30. The normalized relation-

ships are

R( f

, f

, k) = 0.5,

R( f

, f

, k) = 0.75, and

R( f

, f

, k) = 1. We calculate the point S( f

, k) of ﬁle

as follows. Refer to the right-hand side of Figure 6.

The added point S

( f

, f

, k), based on the relation-

ship with f

, is

R( f

, f

, k)· S

( f

, k) = 20·0.75 = 15.

The added point S

( f

, f

, k), based on the relation-

ship with f

, is

R( f

, f

, k) · S

( f

, k) = 30 · 1 = 30.

Therefore, S( f

, k) = 45, namely the sum of the TF–

IDF and both added point components.

For cases involving multiple keywords, the point

is calculated as a sum of the point based on each key-

word. If K is a set of keywords, the point S( f

, K) of

ﬁle f

is calculated as follows.

S( f

, K) =

∑

k∈K

S( f

, k) (7)

4 IMPLEMENTATION

We implemented the proposed method as an experi-

mental system called FRIDAL.

Figure 7 shows the architecture of FRIDAL.

FRIDAL comprises four components, namely the

Web UI, the RDBMS, the Full-text search engine, and

the Controller. The Controller implements two main

functions of our proposed method. First, it mines the

latent interﬁle relationships from the access logs of

Samba, which is widely used as a CIFS (Common In-

ternet File Services) ﬁle server. Second, it performs

the ﬁle point calculations by using latent interﬁle re-

lationships and a full-text search engine.

In an initial setup, administrators provide the ac-

cess log ﬁles of Samba, and FRIDAL starts the min-

ing process with the logs as described in Section 3

and stores relationships to the RDB. When a user re-

trieves ﬁles with FRIDAL, the user ﬁrst supplies the

keywords via the Web UI. The Controller passes the

keywords to Hyper Estraier to obtain the results of a

full-text search. Next, the Controller searches related

ﬁles with latent interﬁle relationships from the RDB

and calculates the ﬁle points. Finally, the Controller

returns the list of ﬁles to the user via the Web UI. The

results comprise the point, the ﬁle path, and basis of

ICSOFT 2010 - 5th International Conference on Software and Data Technologies

240

Table 1: Information about the testers.

recordeing period (days) Total log lines #Files #Relationships Client OS

Tester A 319 4873703 1100 17472 Windows XP

Tester B 319 4323090 713 5692 Windows XP

Tester C 323 7863206 793 5236 Windows Vista

Figure 7: The architecture and behavior of FRIDAL.

the point, such as the TF–IDF or the added points.

We implemented FRIDAL as a web application

written in Java. We used Tomcat 5.5.9 as the web

application container, Oracle Database 10g as the

RDBMS and Hyper Estraier 1.4.13 (Estraier, 2007)

as the Full-text search engine. Since Hyper Estraier

uses an N-gram index, we can avoid Japanese lan-

guage morphological analysis problems, and obtain

a perfect recall ratio by the N-gram method.

FRIDAL derives the ﬁle usage information from

the access logs of Samba 2.2.3a, conﬁgured at debug

level 2. Because we adopted Samba as the ﬁle sys-

tem access logger, we can install FRIDAL with only

slight modiﬁcations to the existing environment. In

addition, we can use historic access log ﬁles even be-

fore installation, because we adopt the default format

for Samba.

5 EVALUATION

We evaluated FRIDAL in two different experiments.

In Experiment 1, we investigate FRIDAL’s ability to

ﬁnd keyword-lacking ﬁles. In Experiment 2, we com-

pare FRIDAL to other search methods with respect to

precision and recall.

5.1 Experimental Environment

We use access logs recording the ﬁle access of three

testers, A, B, and C. Table 1 shows information about

the logs, number of ﬁles appearing in the logs, and the

client OS for each tester. The ﬁle extensions targeted

for analysis are bib, doc, docx, gif, htm, html, jpg,

mpg, mpeg, ppt, pptx, pdf, txt, tex, xls, and xlsx. The

constants described in Section 3.1.1 are T

= 1[m],

= 30[m], T

= 5[h],T

= 10[s]. We specify the pa-

rameters (α, β, γ, δ) = (1, 1, 0.5, 0.5), used in Equation

(3), by performing a preparatory experiment.

5.2 Experiment 1

5.2.1 Experimental Method

In this experiment, we have a tester ﬁnd speciﬁc ﬁles

in another user’s home directory. We examine the

number of times a query is input and ﬁles are checked

when using either FRIDAL or Full-text search. Here,

“query” means keywords plus a ﬁle extension to nar-

row down the search. For search requests, we take

account of the fact that the testers are students and

use the six requests listed in Table 2.

Table 2: The search requests used in Experiment 1.

Search Desired ﬁles

request

The paper of tester A

The source of the image ﬁles in the paper of

tester A

The eight data ﬁles for the paper of tester A

The paper of tester C

The source of the image ﬁles in the paper of

tester C

The data ﬁle for the paper of tester C

5.2.2 Experimental Results

Table 3 gives the experimental results when using the

requests in Table 2. In Table 3, “#Queries” means

the number of times the query was made, “#Check

ﬁles” means the number of the ﬁles checked until the

tester either ﬁnds the ﬁle or gives up the search, and

“Can ﬁnd” means that if the tester can ﬁnd the ﬁle

then “T” is marked, otherwise “F” is marked. If there

SEARCHING KEYWORD-LACKING FILES BASED ON LATENT INTERFILE RELATIONSHIPS

241

Table 3: The cost to search (#Q: Number of Queries, #F:

Number of Checked Files, Ratio: Successful Ratio for ﬁnd-

ing target ﬁles. )

FRIDAL Full-text search

#Q #F Ratio #Q #F Ratio

2 1 1 2 15 1

1 9 1 1 6 0

1 4 3/8 1 0 0/8

1 2 1 1 11 1

1 2 1 1 14 0

1 15 1 2 14 1

are multiple relevant ﬁles, the ratio of the number of

found ﬁles to all ﬁles is given.

5.2.3 Discussion

In Table 3, for Q

, Q

, and Q

, Full-text search can-

not ﬁnd the ﬁles at all, whereas FRIDAL can. With

this result, we can conﬁrm that FRIDAL can ﬁnd

keyword-lacking ﬁles.

For Q

, Q

, whereas both methods can ﬁnd

the ﬁle, the numbers of Queries and Checked ﬁles

of FRIDAL are less than those of Full-text search.

Therefore, it is apparent that FRIDAL can search

more efﬁciently than Full-text search in the case of

searching data related to a paper. Note that, in Q

the tester found the data ﬁle by using Full-text search.

This was because the data ﬁle included the keyword

accidentally.

5.3 Experiment 2

5.3.1 Experimental Method

First, we prepared relevant sets by selecting ﬁles re-

lated to keywords from all ﬁles in the target ﬁle sys-

tem by testers. Table 4 shows the keywords actually

used. Next, we retrieved by those keywords with fol-

lowing four methods.

• FRIDAL

• Full-text search

• Directory search

• Connections calculation

Then we calculate the averaged 11-points precision

of the search results, and also calculate the precision

and recall of the top 20 of the results. If the averaged

11-points precision is high at one recall, the search

result is good because the result will include few mis-

matched ﬁles at that recall.

Directory search is a method that searches the di-

rectory that includes the full-text search result. We

suppose this method is the usual action by users when

ﬁles cannot be found by a full-text search. The re-

sults for directory search are ordered as: the 1st rank

of the full-text search, all ﬁles in the same directory

as the 1st rank of the full-text search, the 2nd rank of

the full-text search, all ﬁles in the same directory as

the 2nd rank of the full-text search (except for ﬁles

already counted), and lower ranks similarly.

Connections calculation is a method that uses

the calculation method of Connections (Soules and

Ganger:, 2005), introduced in Section 2. However

because we cannot obtain the actual ﬁle system calls,

we use the read/write attribute for ﬁle access in the

access logs instead of read()/write(). The constant

parameters in Equations (1) and (2) are set to Time

Window=30[s], P = 2, α = 0.25, and Weight Cut-

off=0.1[%], which are the optimal parameter values

reported in (Soules and Ganger:, 2005).

5.3.2 Experimental Results

Figure 8: The averaged 11-points precision/recall curve of

the results.

Figure 9: The averaged 11-point precision/recall curve of

the results without full-text search results.

Figure 8 is the averaged 11-points precision/recall

graph of the results for six keywords and Figure 9

is the 11-points precision/recall graph of the results

without the full-text search results. The precision at

recall = 0 is the precision at the point when the ﬁrst

relevant ﬁle is found. Table 5 is the average of the

precision and recall of the top 20 of the search results.

5.3.3 Discussion

The precision of FRIDAL is higher than the other

methods at low recalls, as shown in Figure 8 and Fig-

ICSOFT 2010 - 5th International Conference on Software and Data Technologies

242

Table 4: Evaluated query keywords and Relevant ﬁles for Experiment 2.

Keywords Relevant ﬁles #Relevant ﬁles

“Revocation for Encrypted Data Stored with

Replica”

The ﬁles related to the paper written by tester A 203

“DO–VLEI” The ﬁles related to the proposed method named “DO-VLEI” of tester B 206

“DO–VLEI” and “adss” The ﬁles related to the proposed method named “DO-VLEI” of tester B, to the

conference “ADSS”

106

“patent” The ﬁles related to the patent of tester C 48

“Letters” The ﬁles related to paper published in “DBSJ Letters” of tester C 54

“access log” and “ﬁle search” The ﬁles related to the research of tester C 75

Table 5: The averaged precision and recall of the top 20.

Average Average

precision recall

FRIDAL 0.72 0.16

Full-text search 0.62 0.12

Directory search 0.62 0.13

Connections calculation 0.48 0.10

ure 9, and the recall and precision are also better, as

seen from Table 5. These result show that FRIDAL

can retrieve more relevant ﬁles than the others in the

high orders of the results, and so we can ﬁnd the de-

sired ﬁles efﬁciently by using FRIDAL.

We discuss these results in detail in the follow-

ing. In Figure 8, the precision of FRIDAL is slightly

lower than for Directory search and Connection cal-

culation at high recalls. Therefore, the results for

FRIDAL include fewer relevant ﬁles in the low or-

ders of the results. However, because the results of

Directory search and Connections calculation include

vast numbers of ﬁles, the user cannot check the ﬁles

in the lower orders of the results. Thus, the utility

of FRIDAL becomes apparent because the results for

FRIDAL include more relevant ﬁles in the higher or-

ders.

In Figure 8, the reason why the precision of Di-

rector search, at 0.4, is much higher than that of

FRIDAL is that the precision of Director search is

higher than FRIDAL on one keyword. We inves-

tigated this, and found that ﬁles in the relevant set

corresponding to the keyword are organized in just

one directory. For organized ﬁles, searching with

FRIDAL is less efﬁcient than Directory search. Since

we can predict that there are such partially organized

ﬁles, we will improve FRIDAL to prepare for such

cases in future work.

There are two causes for the values of Connec-

tions being lower. First, whereas Connections cal-

culates the point of a ﬁle based on the number of

reading/writing actions for the ﬁle, it does not take

any account of the duration of ﬁle use. Second, it

makes a directed edge from the read ﬁle to the writ-

ten ﬁle, and calculates ﬁle point based on the graph.

The directed edges cannot express interﬁle relation-

ships adequately, because reading/writing ﬁles is not

related to the reference/referenced ﬁle but based on

the application that uses the ﬁle. However, in this ex-

periment, we use the read/write attributes of Samba

instead of the system calls that Connections actually

uses. Therefore, it is not an accurate comparison.

In Table 5, all recalls are of low value because

there are many relevant ﬁles which we cannot ﬁnd re-

lationship between ﬁles include keywords. As shown

in Table 4, there are some correct sets that include

more than 200 ﬁles. However, the number of searched

ﬁles is 20. One of the reasons why we cannot ﬁnd

relationship is that a LaTeX ﬁle of a paper were di-

vided into sub LaTeX ﬁles and each ﬁles had sev-

eral renamed-variant for version control. Since our

method can not deal with ﬁle copy and rename, we

cannot treat access to these ﬁle as ﬁle access to the

same ﬁle. To track ﬁle system changing, such as ﬁle

copy, rename, move, is an issue in the future.

Figure 9 shows how many keyword-lacking ﬁles,

which Full-text search cannot ﬁnd, can be found by

three methods. In these results, the precision of

FRIDAL is also higher than the others. Therefore,

we conclude we will be able to ﬁnd keyword-lacking

ﬁles efﬁciently by using FRIDAL.

6 CONCLUDING REMARKS

Traditional full-text searches cannot search keyword-

lacking ﬁles, even if the ﬁles are related to the key-

words. In this paper, we have proposed a method

for searching keyword-lacking ﬁles, and we have

described a system, FRIDAL, that implements our

method. FRIDAL requires no modiﬁcation in target

systems, and has very small overhead to obtain the in-

formation of latent interﬁle relationships. In addition,

we have performed experiments to evaluate FRIDAL

via testers’ answers. As a result, FRIDAL was found

to be capable of searching for keyword-lacking ﬁles,

with FRIDAL searching more efﬁciently than a full-

text search. In addition, the 11-points precision and

the recall/precision of the top 20 of FRIDAL’s results

were higher than those of the other methods.

In future work, we are considering the following

four points. The ﬁrst is to improve FRIDAL to be

able to deal with the copying, moving, and renam-

SEARCHING KEYWORD-LACKING FILES BASED ON LATENT INTERFILE RELATIONSHIPS

243

ing of ﬁles. Because there are ﬁles whose paths are

recorded in the access logs but that do not exist in

the ﬁle system, it is necessary to deal with the copy-

ing, moving, and renaming of ﬁles, by, for example,

obtaining system calls to ﬁles. The second point con-

cerns the various experiments. Because the experi-

mental target involved the access logs of students, we

only evaluated search requests such as those in Ta-

ble 2. We should perform experiments using a wider

variety of access logs. The third point is to improve

the calculation of the latent interﬁle relationships. Al-

though one relationship is a reference relationship, we

did not take account of the direction of the relation-

ship. Therefore, we should formulate the reference re-

lationship by deriving the direction of the relationship

from the access logs. The fourth point is to improve

the method of displaying the search results. Currently,

FRIDAL displays the ﬁles ordered by ﬁle point on a

line. However, there are cases where the user would

like a graphic structure comprising ﬁles and relation-

ships. Therefore, we should consider ways of display-

ing the search result for all search requests.

ACKNOWLEDGEMENTS

Part of this research was supported by CREST

of JST (Japan Science and Technology Agency),

a Grant-in-Aid for Scientiﬁc Research on Priority

Areas from MEXT of the Japanese Government

(#19024028,#21013017), and the 21

Century COE

Program Framework for Systematization and Appli-

cation of Large-Scale Knowledge Resources.

REFERENCES

Agrawal, N., Bolosky, W. J., Douceur, J. R., and Lorch,

J. R. (2007). A ﬁve-year study of ﬁle-system meta-

data. ACM Transactions on Storage, 3(3).

Barreau, D. and Nardi:, B. A. (1995). Finding and re-

minding – ﬁle organization from the desktop,. ACM

SIGCHI Bulletin, 27(3):39–43.

Blanc-Brude, T. and Scapin, D. L. (2007). What do people

recall about their documents?: Implications for desk-

top search tools. In Proc. Intl’ Conf. on Intelligent

User Interfaces(IUI2007), pages 102–111.

Chirita, P. A., Gaugaz, J., Costache, S., and Nejdl, W.

(2006). Desktop context detection using implicit feed-

back. In Proc. SIGIR 2006 Workshop on Personal In-

formation Management, pages 24–27.

Chirita, P. A., Gavriloaie, R., Ghita, S., Nejdl, W., and Paiu,

R. (2005). Activity based metadata for semantic desk-

top search. In Proc. Second European Semantic Web

Conference(ESWC 2005), pages 439–454.

Chirita, P. A. and Nejdl, W. (2006). Analyzing user be-

havior to rank desktop items. In Proc. Intl’ Symp. on

String Processing and Information Retrieval(SPIRE),

pages 86–97.

Cohen, S., Domshlak, C., and Zwerdling:, N. (2008). On

ranking techniques for desktop search. ACM Transac-

tions on Information Systems, 26.

Dumais, S., Cutrell, E., Cadiz, J. J., Jancke, G., Sarin, R.,

and Robbins, D. C. (2003). Stuff i’ve seen: A system

for personal information retrieval and re-use. In Proc.

SIGIR2003, pages 72–79.

Estraier, H. (2007). Hyper estraier: a full-

text search system for communities.

http://hyperestraier.sourceforge.net/.

Fertig, S., Freeman, E., and Gelernter, D. (1996). ”ﬁnding

and reminding” reconsidered,. ACM SIGCHI Bulletin,

28(1):66–69.

Freeman, E. and Gelernter, D. (1996). Lifestreams: A stor-

age model for personal data. ACM SIGMOD Bulletin,

25:80–86.

Gifford, D. K., Jouvelot, P., Sheldon, M. A., and James

W. O’Toole, J. (1991). Semantic ﬁle systems. In Proc.

ACM Symposium on Operating Systems Principles,,

pages 16–25.

Google (2010). Google image search,. http://images.

google.com/.

Hayes, B. (2002). Terabyte territory. American Scientist,

90(3):212–216.

Ishikawa, K., Morishima, A., and Tajima:, K. (2006). De-

velopment of a semantic ﬁle system for managing

large document spaces(in japanese). Technical report

DE2006-115, IEICE.

Namazu (2009). Namazu: a full-text search engine.

http://www.namazu.org/index.html.en.

Nejd, W. and Paiu, R. (2005). Desktop search – how con-

textual information inﬂuences search results and rank-

ings. In Proc. ACM SIGIR 2005 Workshop on Infor-

mation Retrieval in Context (IRiX), pages 29–32.

Ohsawa, R., Takashio, K., and Tokuda, H. (2006). Oredesk:

A tool for retrieving data history based on user opera-

tions. In Proc. Eighth IEEE International Symposium

on Multimedia (ISM’06), pages 762–765.

Page, L., Brin, S., Motwani, R., and Winograd, T. (1999).

The pagerank citation ranking: Bringing order to the

web. TR 1999-66, Stanford InfoLab.

Rekimoto, J. (1999). Timemachine computing: A time-

centric approach for the information environment. In

Proc. ACM UIST’99.

Soules, C. A. and Ganger:, G. R. (2005). Connections: Us-

ing context to enhance ﬁle search,. In Proc. ACM Sym-

posium on Operating Systems Principles, pages 119–

132.

Watanabe, T., Kobayashi, T., and Yokota:, H. (2008). A

method for searching keyword-lacking ﬁles based on

interﬁle relationships. In Proc. 16th Intl’ Conf. on

Cooperative Information Systems (CoopIS’08), pages

14–15.

Yee, K.-P., Swearingen, K., Li, K., and Hearst:, M. (2003).

Faceted metadata for image search and browsing,. In

Proc. CHI’03, pages 401–408.

ICSOFT 2010 - 5th International Conference on Software and Data Technologies

244