Professional Documents
Culture Documents
net/publication/261149618
CITATIONS READS
10 337
3 authors, including:
Zhifang Liao
Central South University
42 PUBLICATIONS 164 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Zhifang Liao on 09 March 2019.
Abstract—In this paper, we combine probabilistic model and similarity of user profile and each profile of all the items and
classical content-based filtering recommendation algorithms to recommends items may satisfy user need or tastes.
propose a new algorithm for recommendation system, which we Combining the feature of items and the user interest model,
call content-based filtering recommendation algorithm using the utility function is usually defined as:
HMM. We utilize the HMM of recommended items to match user
model and recommend items using user data. According to u (c, s ) = score(ContentBaseProfile(c), Content ( s))
experiment result, this new method is more effective on
describing a user’s interest compared with the VSM-based
algorithm. There are many methods to calculate the utility, for
example, the cosine similarity measure:
Keywords-HMM; content-based filtering; recommendation system JJG JJJG
JJJG JJG w ⋅w s
I. INTRODUCTION u (c , s ) = cos( w c , w s ) = JJJG c JJG
The rapid development of internet brings the problem of w c × ws
information overload: user receives too much potential
¦
valuable information to approach the most useful part, which k
causes the reducing of efficiency [1]. Recently, many wi , c wi , s
applications of network (web portal, search engine, data index, = i =1
¦ ¦
k k
etc.) help users filter the information. However, all of these
methods just satisfy the basic needs and do not consider the
( i =1
w i2, c i =1
w i2, s )
personalized services to solve the information overload
problem. As an important method of information filter, Finally, the system sorts items by the utility values and
recommendation system provides personalized services to recommend other items to that user which have a high degree
users through identifying and predicting user preferences and of similarity to the user’s profile.
become a most useful method helping user to deal with III. HIDDEN MARKOV MODEL
information overload [2].
The most typical application of recommendation system is The basic theory of hidden Markov model was published
in the e-commerce which has a promising development in a series of classic papers by Baum and his colleagues in the
prospect. The online mall can recommend customers products, late 1960s and early 1970s and was firstly applied in speech
which may attract their attention or satisfy their needs (such as processing applications in the middle of 1970s [4]. From then
books, video, etc.), according to the users’ interests. It can on, HMM has been widely implemented in many fields.
enhance the products sales by meeting the potential needs of HMM is a probabilistic statistical model. A discrete hidden
users, which are usually not clear and fuzzy. Recommendation Markov model contains a set of states and an alphabet of
systems have become extremely common in recent years. output symbols [4]. States can both transform from one to the
Almost all of the large e-business systems have applied other and emit several output symbols. The probabilities of the
recommendation system in different ways, such as Amazon, two procedures make up the transition distributions over states
eBay, etc. and the emission distributions over the output symbols
respectively.
II. CONTENT-BASED FILTERING ALGORITHM A discrete hidden Markov model can be characteristically
The content-based recommendation algorithms derive defined as:
from information retrieval and filtering research [3]. The
content-based recommendation is the continuation and • A set of hidden states Q = {Q , Q ,!, Q } , where
development of early collaborative filtering method. Content- 1 2 N
based recommendation systems recommend items similar to Q (1 < t < N ) is the state in time t.
those that the user has selected in the past rather than the user’s t
comments of items. Many current content-based systems build • A set of output symbols X = { X , X ,! , X }
the users’ profile and items’ profile. A user profile contains 1 2 M
information about user’s tastes, preferences and needs which emitted by each hidden states.
can be elicited from users’ questionnaires or learned from their
transactional behavior over time, while an item profile contains • The probabilities distribution of state transitions:
a set of attributes of items. Then the system calculates the
276
We compare the precision and recall rates of vector space However, it is difficulty to apply an all-purpose
model based algorithm (VSM-Based algorithm) and HMM- recommendation algorithm and describe user interests precisely
based algorithm and Figure II shows the experiment result. for different background. In future work, we will focus on the
According to the result, we can see the average precision of development of relational techniques and keep improving the
HMM-based algorithm is higher than VSM-Based algorithm. average precision of our algorithm.
The main reason is that VSM-Based algorithm calls for
accurate match, however, the number of similar key words REFERENCE
which are shared between items and user interests is so few [1] D. Bawden, C. Holtham, and N. Courtney. “Perspectives on information
generally that the precision of VSM-Based algorithm overload,” Aslib Proceedings , vol. 51, Sep. 1999, pp. 249-255.
decreases sharply. By contrast, HMM-based algorithm can [2] G. Adomavicius and A. Tuzhilin. “Toward the Next Generation of
avoid this problem by utilizing probabilities distribution to Recommender Systems: A Survey of the State-of-the-Art and Possible
Extensions,” IEEE Transactions on Knowledge and Data Engineering,
measure similarity between items and user interests. Therefore, vol. 17, Jun. 2005, pp. 734-749.
the average precision of HMM-based algorithm is higher. [3] R. Baeza-Yates and B. Ribeiro-Neto. “Modem Information Retrieval,”
New York: Addison-Wesley Publishing Co, 1999, pp. 271–350.
VI. CONCLUSION
[4] L. R. Rabiner, “A Tutorial on Hidden Markov Models and Selected
We propose content-based filtering recommendation Applications in Speech Recognition,” Proceedings of the IEEE, vol. 77,
algorithm using HMM for personalized recommendation Feb. 1989, pp. 257–286.
system. Compared with VSM-Based algorithm, the new [5] S. Zhong and J. Ghosh. “A Unified Framework for Model-based
method is more effective on describing a user’s interest. Clustering,” Journal of Machine Learning Research, vol. 4, Dec. 2003,
pp.1001-1037.
Figure I
5HVXOWV&RPSDUDWLRQ
3UHFLVLRQ
5HFDOO5DWH
960%DVHG +00%DVHG
Figure II
277