User Profile Generation Based on a Memory Retrieval
Theory
Fabio Gasparetti and Alessandro Micarelli
Universita degli Studi Roma Tre,
Dipartimento di Informatica e Automazione
Via della Vasca Navale, 79 00146 Roma, Italy
Abstract. Several statistical approaches for user profiling have been proposed in
order to recognize users’ information needs during their interaction with informa-
tion sources. Human memory processes in terms of learning and retrieval are with
no doubt some of the fundamental elements that take part during this interaction,
but actually only a few models for user profiling have been devised considering
explicitly these processes. For this reason, a new approach for user modeling is
proposed and evaluated. The grounds in the well-studied Search of Associative
Memory (SAM) model provides a clear definition of the structure to store infor-
mation and the processes of learning and retrieval. These assets are missing in
other works based for example on simplified versions of semantic memory mod-
els and co-occurrence data.
1 Introduction
In a large repository of information, such as the Web, the importance of personalization
is certain. The instant availability of many interesting resources is a great opportunity
that users can exploit for many of their every day tasks. Nevertheless, when the amount
of information exceeds the time a user spends to read and understand it, information
overload issues have to be addressed. For this reason, the identification of the user infor-
mation needs and the personalization of the human-computer interaction are becoming
important research topics in this field.
There are a couple of important advantages if we choose to ground the personaliza-
tion’s user profiling on cognitive theories and models. Decades of studies in this field
are available, each of them investigated through several kinds of evaluations. Further-
more, learning processes and memory structures could be considered as the foundation
level where we can start building more sophisticated information behavior and seeking
model instances.
Two systems use natural language processing and semantic or keyword networks
in order to build long term user profiles and evaluate the relevance of text documents
with respect to a profile: the SiteIF project [1] and the ifWeb prototype [2]. Even if
that approach is sometimes called cognitive filtering, beside an implicit reference to
simplified versions of the Quillian’s semantic knowledge model, these two systems do
not represent user needs with structures that are directly inspired by cognitive theories
and models.
Gasparetti F. and Micarelli A. (2005).
User Profile Generation Based on a Memory Retrieval Theory.
In Proceedings of the 1st International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces, pages 59-68
DOI: 10.5220/0001422500590068
Copyright
c
SciTePress
Two other works based on cognitive research but more focused on the prediction of
user actions are: the SNIF-ACT model [3], and the Cognitive Walkthrough for the Web
(CWW) [4].
In this work, we describe an approach to build user models that aims to emulate
the process of learning and recovering of information that occurs in human memory.
This is one of the innovative aspects compared with more traditional approaches, where
the structure and the items represented are the major research subjects. We narrow the
approach to the human-computer interaction in information environments where the
user usually analyzes and selects documents in order to satisfy given information needs.
In the next section, we will briefly describe SAM, the general theory on which we
have based our approach that is the subject of the following section along with all the
advantages and possible issues. At the end, an instance of this model will be applied
and evaluated in the context of the Internet browsing activities.
2 User Profiling
Before investigating in details the proposed user modeling approach, a brief introduc-
tion of the SAM theory is given. For a closer examination of this theory see for example
[5].
2.1 SAM: Search of Associative Memory
The SAM is a general theory of retrieval from long-term memory that considers both
the structure of the memory system and the processes operating within it. The struc-
ture refers to the items represented and their organization in the memory system, the
processes refer to the main activities that occur within the memory, such as learning
and retrieval.
The organization of the memory is divided in two parts: the Long-Term Store (LTS)
and the Short-Term Store (STS). The STS shows two key features: it has a limited ca-
pacity and it is easily prone to forget its content. It can be seen as a temporarily activated
subset of information enclosed in the permanent storage LTS. Its role corresponds to a
working space for control processes, such as coding, rehearsal, decisions, etc.. When a
new sensory input occurs, the related information is analyzed through the LTS structure.
The result of this analysis is the activations of some information units that are placed in
the STS. It is also possible to update the STS by means of internally generated probe
cues, composed of information previously retrieved from the LTS and currently still in
the STS.
Both kinds of memories are composed of unitized images, the objects that may by
sampled and recovered during a memory search. The retrieval process is based on the as-
sociative relationships between probe cues and memory images that can be represented
by means of a matrix of retrieval strengths. These strengths determine the probability
for a given set of probe cues to gather an image stored in the LTS. Basically, the user
model is built by means of all the strengths stored in that matrix.
In the retrieval process, along with the current cues, it is also possible to consider
the contextual information related to the current user’s task, such as, temporal data,
60
Fig.1. An extension of the retrieval process flowchart in Raaijmakers and Shiffrin 1981 to con-
sider the external cues originated by the interaction of the user with the information sources and
theirs effect on the process. The broken box margins the original chart.
category and item names, etc.. Actually, the current cues used during the learning and
retrieval can be seen as a context as well. If we store an image associated with the cur-
rent cues, some relationships between this image and the cues are created, or increased
if they already exist. Each time we have the same cues as context, such an image will
tend to be retrieved. As we will describe in the next section, the context-image relation-
ship is viewed as an essential feature in the user modeling approach.
The LTS is probed with the current information related to the user’s task, the in-
formation available in the STS and the one retrieved earlier in the search. This process
determines what image is sampled and made available to the user for evaluation and
decision-making. It is convenient to separate the sampling phase in two parts: (1) Be-
61
cause just a small number of images in the LTS have a non negligible strength to the
current probe cues, an initial set is built including those images in LTS, (2) After the
sampling phase, the image elements that will be activated and made available to the user
is drawn. This recovery process depends on the strength between the selected images
and the probe cues as well.
For the building and updating of the LTS structure, it is possible to use the same
storage approach followed in the SAM Simulation (SAMS). This approach consists of
a buffer rehearsal process [6] that updates the item-item and context-item strengths as
a function of the total amount of time the pair’s objects are simultaneously present in
the rehearsal buffer. This buffer corresponds to the STS, so it has a limited size and
stores all the cues plus the current context. If the buffer size is reached, an item will be
randomly replaced.
The computational details of the theory, along with all the adaptations we made for
the context under examination will be discussed in the next section.
As the authors of the SAM theory have emphasized, the long-term memory images
and probe cues are quite distinct. The images in memory correspond with the past cues
plus the context at the moment of the learning. This relationship is used in the sampling
process to identify the images related to a given context, but also to be able to retrieve
all the connected cues stored during the learning.
2.2 The Proposed User Modeling Approach
During each information seeking session where the user is looking for sources able
to satisfy particular information needs, the user must assemble and deal with a set of
concepts related to these needs. If we were able to identify all these concepts it would be
possible to personalize the interaction with the sources and, at the same time, filtering
the information that could be retrieved by autonomous search tools. Because human
memory is involved each time the user runs into a new interesting concept, whether he
has to learn or retrieve it from his memory, a technique that aims to emulate this memory
could be successfully used as user modeling component in a personalized system.
Unfortunately, the user does not usually confine the interaction to one or few in-
formation sources. Talking to colleagues, friends and relatives, reading newspapers and
magazines, watching TV are just examples of the possible interactions a user can be
involved in. Nevertheless, we assume that given a information need, if we look at the
different kinds of interactions, the user will not ignore the most important concepts.
Therefore, if we restrict our analysis to one of the possible interactions, it is always
possible to recognize the subset of the most valuable concepts.
Although the SAM theory does not require a particular memory representation, for
the task under consideration we need to define how the information is encoded. In this
study we have chosen the word as the atomic unit of information that is possible to store
in the LTS and STS. This choice is primarily motivated by its simplicity and adaptation
to the personalization context. We are considering environments where the information
is represented by text sentences, so the word seems the obvious element that could be
stored and analyzed.
Natural language processing techniques could be also used to raise the representa-
tion level of the units. If we were able to identify the concept related to a particular word
62
or sentence, we could actually store this information in the memory structure, getting
closer to a semantic network definition. However, due to the low reliability of these
techniques, in this study we have decided to restrict the representation to the word.
As we have explained in the previous section, the retrieval process consists of two
steps: the sampling of the images in the search-set, and the recovery of the information
contained in the sampled image. Both steps make use of a strength matrix, where the
generic value S
T
(Q
ij
, Ij) corresponds to the strength between a cue Q
ij
and the image
I
j
. The rows Q
ij
refer to the images, which in our case are words, I
j
and the Q
i0
values
refer to the context cue.
The first step draws all the sampling probabilities of sampling a given image I
i
in
the LTS as a function of the current cues Qs:
P
S
(I
i
|Q
1
, Q
2
. . . Q
M
) =
Q
M
j=1
S
T
(Q
j
, I
i
)
W
j
P
N
k=1
Q
M
j=1
S
T
(Q
j
, I
k
)
W
j
(1)
where M is the number of cues, and N is the matrix dimension. The W
j
values are
weights used to set the importance of the single cues. In our domain for example, these
weights could be set as a function of the Inverse Document Frequency (IDF) values in
order to decrease the weights of the words that frequently occur in a given corpus and
that can be considered very common and therefore little relevant.
This formula gives the highest probability to the terms with the highest product of
strengths, and hence those that tend to be greatly associated to all the current cues.
Once a word has been sampled, the recovery process step takes place. For the sam-
pled image I
i
we draw the probability:
P
R
(I
i
|Q
1
, Q
2
. . . Q
M
) =
P
M
j=1
S
T
(Q
j
, I
i
)
W
j
P
N
k=1
P
M
j=1
S
T
(Q
j
, I
k
)
W
j
(2)
The currently activated information in the STS and its relationship with the one in
the LTS is used to build new relations and structures in the memory. The important
feature of the theory kept in our approach is that the information currently active in the
STS tends to be stored together. So in each storage phase, an item is stored along with
all the contextual information, the same information that could help retrieving the item
later, even if the STS misses a direct reference to it.
In natural language domain, such as the one under consideration, the context helps
also to disambiguate the meaning of a word stored in the LTS. The relationships among
items bind each word with the context at the time of learning. If the user runs into the
same word but with different contexts, different sets of relationships will be created
between the item and the contexts. During the retrieval, if we probe just the ambiguous
item it is not possible to have any information that helps us to recognize its meaning.
But if we add just one element contained in one of the contexts, the retrieval process
has much more chance to return other items related to the context at issue.
Moreover, the importance of one term for the user is not drawn just by a single value
or weight associated to it, as it happens in the Vector space model or in many Bayesian
approaches. A term is judged interesting if it is also bound to a context that is related to
63
the current context. In other words, the importance is not longer an becomes an absolute
value but a relative one.
The learning process that corresponds to the process described in the SAM theory
updates the strengths in the LTS as a function of the total time a word or a pair of words
is stored together in the STS buffer. Given t
i
the time spent in the buffer by the generic
image, i.e., word, I
i
, and given t
ij
the time the images I
i
and I
j
occur together in the
buffer, we have:
S
T
(C, I
i
) = at
i
S
T
(I
i
, I
j
) = S
T
(I
j
, I
i
) = bt
ij
, t
ij
6= 0
S
T
(I
i
, I
i
) = ct
i
If a word pair have never appeared together in the buffer, they assume anyway a
non negligible residual retrieval strength d. The values a, b and c are parameters of the
model, and C is the context currently in the STS.
The strengths are also updated each time a particular combination of cues is able to
sample and recovery a particular image. In this case the strength between the cues and
the sampled image, and the self-association strength is incremented:
S
′
T
(C, I
i
) = S
T
(C, I
i
) + e
S
′
T
(I
i
, I
j
) = S
′
T
(I
j
, I
i
) = S
′
T
(I
j
, I
i
) + f, t
ij
6= 0
S
′
T
(I
i
, I
i
) = S
T
(I
i
, I
i
) + g
where the S are the strengths before the incrementing and e, f and g are other
parameters.
Now that we have explained how the user model is built and updated, the general
retrieval process in Fig.1 that occurs each time the user interacts with a information
source is described. This process aims to simulate as much as possible the analysis
and recovery of the concepts that the user goes through during the information seeking
process.
The process is composed of the inner retrieval process that resumes the one dis-
cussed in the SAM theory. In the first cycle there are at the most K
MAX
attempts to
sample and recovery a word stored in the LTS by means of the current context. A fail-
ure is every attempt that does not lead to recall of a new item. This may happen either
because of low recall probabilities, Eq. 2, or if the image has been already unsuccessful
sampled in a previous cycle and the context has not been altered meanwhile.
If a word has been successfully sampled, the strengths to the current context S
T
(C, I
i
)
and the self-association S
T
(I
i
, I
i
) are incremented, and the word is added to the context
which is used in the next cycle to sample other correlated images. If the recovery fails,
than the counter K and L are incremented to see if we are at end of the cycle. Each time
a word is sampled in the second cycle, it joins the context and the strengths S
T
(C, I
i
),
S
T
(I
j
, I
j
) and S
T
(I
i
, I
j
) are updated.
The outer elements of the flowchart concern the information that comes from the in-
teraction between the user and the information sources. Here we consider each data that
has some kind of relation with the user’s information needs and that can be represented
in a text format. Possible examples are: queries, document’s snippets or categories’
keywords selected by the user.
When new information is available it is added to the context and a new retrieval
process takes place to recovery all the correlated information and, at the same time,
64
store the data in the LTS. At the end of the retrieval process, we check if the user has
completed his seeking session, in this case we leave the main cycle of the process. If the
user is still trying to satisfy his information needs, the old context is compared to the
current one in case the user has changed the domain in which he is working on. A new
context means a reset of the STS, that is, all the temporary information stored during
the session up to the context change is wiped out.
Each time information is available and the recovery process is completed, we are
able to collect the information stored in the STS that can be used to give a representation
of the current information needs of the user. This representation regards the last context
that has been identified during the search.
It is possible to see how the approach can work without the user’s explicit feedback,
especially if the cues originated from the user interaction with the information sources
are fairly representative. Beside taking time during the seeking processes, past studies
showed how explicit feedbacks are not able to considerably improve the user model
especially if a good interface to manage the model is not provided [7].
A last remark on this approach is the absence of the renting technique often used
in order to remove concepts from the user model that are no longer judged interesting
for the user. There are two reasons to justify the presence of this technique. During
the information behavior the user learns concepts that influence his knowledge and
therefore his representation of the information needs. A user model should be able
to recognize these knowledge alterations and their influences on the user’s goal. The
second reason regards the accuracy of user models. Many content-based techniques
take a representation of the documents’ content seen by the user as the primary source to
build the model. But many times documents concern with different topics, and the user
is interested in just a subset of them. If we update the model with irrelevant information,
a technique to forget some concepts no longer considered is needed in order to improve
the model as time goes by.
One of the interesting features of the discussed approach is that the model is af-
fected just by an additive learning process. It does not mean that it is impossible to
ignore a concept that is judged not interesting. Even if no deletions can occur, the learn-
ing process tends to increase the strengths of the relationships of concepts frequently
covered during the seeking process. A wrong concept occurred during the process is
stored in the LTS but its strengths with other images will not be intensified in the future.
As a result, the probability to recovery this item given a context will get low values. In
other words, forgetting is a result of the failure in the attempt to retrieve a concept.
2.3 An User Model Instance for Browsing Activities
In this section, we want to briefly describe an instance of the proposed user profiling ap-
proach in a concrete domain, as the Internet browsing. This is a very important domain
because of the amount of the available data that can be exploited during any informa-
tion seeking task. That is one of the reason the personalization of the interaction with
the search tools is becoming an important research topic. Identifying what a user is cur-
rently looking for during a browsing session, that is his information needs, is the first
step toward an efficient personalization.
65
In order to have an instance of the user modeling approach, a methodology to iden-
tify the cues that will be used in the learning and storage process discussed in the pre-
vious section is needed. The notion of Information scent [8,9] developed in the context
of Information foraging theory [10], is a valid choice to identify these cues by means
of the text snippets associated to the links. Users use these browsing proximal cues to
decide if access to the distal content, that is, the page at the other end of the link. For-
mally, the information Scent is the imperfect, subjective perception of the value or cost
of the information sources obtained from proximal cues.
Therefore, it is possible to begin a learning and retrieval process each time a user
selects a link and visit a page, considering the anchor texts of the chosen link as the
cue. But from preliminary evaluations we have verified that this information could be
not enough to recognize valuable cues for the learning process, especially if the text
consists just of few words with no correlation, e.g., “full story”, “page two”, “link”,
“rights reserved”, etc..
For this reason we have developed an algorithm whose goal is to collect all the
information related to a given link selection. In a few words, the text of a link is joined
with the title of the linked page that has been visited by the user. Then the context of a
link, i.e., the text surrounding the anchor is retrieved by means of the page’s Document
Object Model tree representation. In other words, the page is divided in units whose
boundaries are arranged by a subset of HTML tags, e.g., TR, P, UL, etc., and the text of
the deepest unit that contains the link is set as the context of the link. At the end, this
context is compared to all the other units of the page that includes the link, in order to
find further related text for the cue. The comparison is based on the same IR similarity
function used in the TextTiling algorithm [11].
For simplicity, in this evaluation we have ignored the explicit representation of con-
texts. Actually a context is an important facet of the approach because it let the model
identify multiple topics and create for each topic particular relationships with the im-
ages, i.e., words, unlike many other modeling techniques where each word is evaluated
as important or not regardless of the current context. But it is however possible to obtain
interesting results taking under consideration just the implicitly relationships between a
word and the context represented with the other words currently stored in the STS. This
alteration affects the retrieval process replacing the S
T
(C, I
i
) with the set of S
T
(I
j
, I
i
)
for all the I
j
images currently in buffer.
3 Evaluation
In order to evaluate the proposed user modeling approach, we have collected a set of
browser histories from a certain number of users. After a filtering process needed to
remove not interesting Web sites, e.g., Web mail services, we have selected browsing
subsequences related to topics that we were able to easily recognize. An external per-
son selected a subset of these subsequences and, therefore, a subset of the topics. The
resulting subsequence set has been the input of our user model.
A traditional user model has been employed to provide a benchmark to compare our
results. It is based on a Vector Space model (VSM). A Relevance Feedback (RF) tech-
nique is used to update the model with the information collected during the browsing.
66
The input of the two user models is the information collected by the technique de-
scribed in the previous section based on the anchor texts, the titles and the correlated
text in a Web page. The images stored in the user model correspond to the word ex-
tracted from this information. A more refined approach, where an image is associated
with the meaning of the word has not yet been considered.
During the evaluation, we have submitted the selected subsequences. At the end we
have collected 25 keywords with the highest score from the VSM-based user model,
i.e., the highest frequency of occurrence. For the proposed user model, we have instan-
tiated the recovery process described in Sec. 2.2, see Eq. 2. In this way we are able to
collect the same number of keywords related to the given probe cue, as a function of
the information stored in the model.
An external judge assigned a score in 4 point relevance Likert scale for each key-
word extracted from the user models, from no-relevance to high relevant, according to
the topic under consideration.
The parameters of the model get these values: a = 0.1, b = 0.1, c = 0.1, d = 0.2,
e = 0.7, f = 0.7 and g = 0.7. The maximum dimension of the rehearsal buffer is 100
elements.
Table 1. Evaluation results for the seven sequences of browsing histories analyzed building a
single user model.
Topic VSM +
RF
Proposed Ap-
proach w/Input
New England Execution 0.16 0.8
Green Day Band 0.2 0.16
Podcaster 0 0.4
Tsunami 0 0
Ray Charles 0.24 0.76
Popular T-Shirt 0.16 0.8
Vegetarian Diet 0.16 0.2
The results are shown in Table 1. It is possible to note how the proposed user model
outperforms the traditional VSM approach. This phenomenon is related to inability of
the VSM-based user model to store browsing contexts. The evaluation shows how the
instantiation of the proposed user modeling approach is able to recognize important
keywords as a function of the current context, i.e., the probe cues, in comparison with
a traditional user model based on the Vector Space model and the Relevance Feedback
technique. Once we are able to identify those keywords, it is possible to use that model
for example in the filtering task, re-ranking the results of a search engine by means
of the keywords returned from the user model after a browsing session, or building
hypertext pages adapted to the user needs.
67
4 Conclusion
Beside describing an approach to build user profiles, which aim to recognize concepts
related to the current information needs, this work suggests a standpoint where the
user profiles are seen consisting of multiple layers. The low level regards the basic
human memory processes of learning and retrieval while the higher levels deal with the
information seeking strategies a user can undertake in order to reach a particular task.
An implementation of a user model based on the proposed approach has been briefly
showed along with an evaluation that has produced interesting results. The future work
will concern the extension of the approach to include other information that can be
collected, such as the previous contexts occurred during the learning or the category of
the current images. The latter information can be easily obtained by one of the common
categorization techniques in the literature.
References
1. Magnini, B., Strapparava, C.: User modelling for news web sites with word sense based
techniques. User Modeling and User-Adapted Interaction 14 (2004) 239–257
2. Asnicar, F.A., Tasso, C.: ifweb: a prototype of user model-based intelligent agent for docu-
ment filtering and navigation in the world wide web. In: Proceedings of Workshop Adaptive
Systems and User Modeling on the World Wide Web (UM97), Sardinia, Italy (1997) 3–12
3. Pirolli, P., Fu, W.T.: Snif-act: A model of information foraging on the world wide web. In:
Proceedings of the 9th International Conference on User Modeling, Johnstown, PA, USA
(2003)
4. Blackmon, M.H., Polson, P.G., Kitajima, M., Lewis, C.: Cognitive walkthrough for the web.
In: Proceedings of the ACM conference on human factors in computing systems (CHI2002),
Minneapolis, Minnesota, USA (2003) 463–470
5. Raaijmakers, J.G., Shiffrin, R.M.: Search of associative memory. Psychological Review 88
(1981) 93–134
6. Atkinson, R.C., Shiffrin, R.M.: Human memory: A proposed system and its control
processes. In Spence, K., Spence, J., eds.: The psychology of learning and motivation: Ad-
vances in research and theory. Volume 2. Academic Press, New York, USA (1968) 89–195
7. Wærn, A.: User involvement in automatic filtering: An experimental study. User Modeling
and User-Adapted Interaction 14 (2004) 201–237
8. Chi, E.H., Pirolli, P., Pitkow, J.: The scent of a site: A system for analyzing and predicting
information scent, usage, and usability of a web site. In: Proceedings of the Conference on
Human Factors in Computing Systems (CHI2000), Hague, Netherlands (2000) 161–168
9. Chi, E.H., Pirolli, P., Chen, K., Pitkow, J.: Using information scent to model user information
needs and actions on the web. In: Proceedings of the ACM Conference on Human Factors in
Computing Systems (CHI2001), Seattle, WA, USA (2001) 490–497
10. Pirolli, P., Card, S.K.: Information foraging. Psychological Review 106 (1999) 643–675
11. Hearst, M.A.: Texttiling: Segmenting text into multi-paragraph subtopic passages. Compu-
tational Linguistics 23 (1997) 33–64
68
