Ymal : ¸íá Óýóçìá ÓõóÜóåùí Ó÷åóéáêþí ÂÜóåùí

ÄåäïìÝíùí

1 2 1
Efty hia Koletsou , Kostas Stefanidis , Marina Drosou , Evaggelia Pitoura
1

1
Dept. of Computer S ien e, University of Ioannina, Gree e. {ekoletso, mdrosou, pitoura} s.uoi.gr
2
Dept. of Computer S ien e and Engineering, Chinese University of Hong Kong, Hong Kong. kstef uhk.edu.hk

Abstra t| Óå áõÞí çí åðßäåéîç ëïãéóìéêïý ðáñïõóéÜ- returns a set of results Res(Q) in the form of tuples, possibly
æïõìå éò Ymal óõóÜóåéò, Ýíá ðëáßóéï åñãáóßáò ðïõ åðåêåßíåé produ ed by joining several relations of D. Besides Res(Q), we
á ó÷åóéáêÜ óõóÞìáá âÜóåùí äåäïìÝíùí ìå ëåéïõñãéêüçá would like to lo ate and re ommend to users a set of tuples
óõóÜóåùí. ÓõãêåêñéìÝíá, ðñïåßíïõìå, ìáæß ìå á áðïåëÝ-
óìáá ìßáò åñþçóçò, íá óõóÞíïõìå óï ÷ñÞóç åðéðëÝïí that may also be of interest to them. We all this set of tuples
áðïåëÝóìáá ðïõ ïíïìÜæïõìå \You May Also Like" Þ Ymal \You May Also Like" tuples or Ymal results. We denote this
áðïåëÝóìáá. éá íá õðïëïãßóïõìå á Ymal áðïåëÝóìáá set as Ymal(Q).
ãéá ìßá óõãêåêñéìÝíç åñþçóç, ÷ñçóéìïðïéïýìå åßå ìüíï ï To ompute Ymal re ommendations, we exploit the ontent
ðåñéå÷üìåíï ïõ áðïåëÝóìáïò çò åñþçóçò (ïðéêÞ ðñïóÝã-
ãéóç) åßå êáé ï ðåñéå÷üìåíï çò âÜóçò äåäïìÝíùí (êáèïëéêÞ and s hema of the urrent query result and database instan e.
ðñïóÝããéóç). We onsider an SPJ query Q of the form:
sele t A from R where P1 AND P2,
I. Introdu tion
R denotes a set of relations of D, A a set of attributes
The typi al intera tion of a user with a database system is {A1 ; : : : ; An } of the relations of D, P1 the onjun tion of the
by formulating queries. This intera tion mode assumes that join onditions for Q and P2 the onjun tion of the remaining
users are to some extent familiar with the ontent of the sele tion onditions for Q. For pro essing re ommendation
database and also have a lear understanding of their infor- requests, we re-write the submitted query Q into a set of queries
mation needs. However, as databases get larger and a essible referred to as Ymal queries. An Ymal query is of the form:
to a more diverse and less te hni ally-oriented audien e, a new sele t B from S where P.
\re ommendation"-oriented form of intera tion seems attra - In the following, we will show our query re-writing me hanism
tive and useful. for omputing B, S, and P. This me hanism is based on either
In this paper, motivated by the way re ommenders work, (i) lo al analysis of the intrinsi properties of the result Res(Q)
we onsider \re ommending" to the users tuples not in the or (ii) global analysis of the properties of the database D.
results of their queries but of potential interest. For instan e, The union of the results of Ymal queries onstitute the Ymal
when asking for movies with dete tives, we ould re ommend re ommendations, or Ymal(Q).
movies with poli emen as well. When looking for drama movies Next, we des ribe the main omponents of the ar hite ture
produ ed in England with Os ar nominations, we ould also of our system. A high level representation is depi ted in Fig. 1.
re ommend similar movies with BAFTA awards. We all su h On e a query Q is submitted, we ompute its a tual result set.
results \You May Also Like" or Ymal results for short. Ymal Using the query results, we onstru t a set of Ymal queries
results are useful be ause they let users see other tuples in the by re-writing Q. The results of these queries orrespond to the
database that they may be unaware of. Ymal results.
Extending database queries with re ommendations has also Attribute Sele tion Generator. This omponent takes as input
been suggested in two very re ent works, namely [2℄ and [1℄. the submitted query Q and returns the attributes B that will ap-
[2℄ proposes a general framework and a related engine for the pear in the sele t lause of the Ymal queries. In a lo al-based
de larative spe i ation of the re ommendation pro ess, while approa h, where only Res(Q) is employed, all Ymal queries
here, instead, we propose a spe i re ommendation method for have the same sele t lause whi h ontains all attributes B
relational databases. Re ommendations in [1℄ are based on the that appear in the onditions P2 of the initial query.
past behavior of similar users whereas we onsider the ontent
of the database. A preliminary version of our re ommendation In a global-based approa h, where the ontent of the
fun tionality is presented in [3℄. database is also taken into a ount, we use a spe ial sele t
In this demonstration paper, we brie y des ribe Ymal re - lause. We simply onstru t this lause by removing from the
ommendations for relational databases and present the ar hi- set of attributes appearing in the relations of the from lause
te ture of our system for pro essing re ommendation requests of the under- onstru tion Ymal query, the non-informative
(Se tion II). We also des ribe our demonstration of Ymal attributes, su h as the primary and foreign key attributes
re ommendations for a relational movies database system (Se - ontaining meaningless values.
tion III). Relation Sele tion Generator. This omponent takes as input
the query Q and returns the relations S that will appear in the
II. Ymal Re ommendation System Overview
from lause of the Ymal queries. In the lo al-based approa h,
Assume a relational database system D and a set of users S = R, while in the global-based approa h, S is a superset of
intera ting with it by posing traditional sele t-proje t-join R. To ompute S, we maintain orrelations among relations. A
(SPJ) queries. Given a query Q, a typi al database system relation Ri is orrelated to a relation Rj with a s ore p that
 Fig. 1. Ymal system ar hite ture.
re e ts the result size of their join. This way, we onstru t S by
adding to R the k most orrelated relations to those in R, where
k is determined by an input ardinality onstraint. The higher
the s ore p of a relation, the more orrelated the relation.
Constraint Sele tion Generator. This omponent generates the
sele tion onditions that the Ymal queries ontain taking as
input both Q and Res(Q). Again, we distinguish between two
ases; sele tion onditions are lo ated either with respe t to a
lo al analysis of the intrinsi properties of the result Res(Q),
or based on a global analysis of the properties of the database
D.
In the former ase, given a query Q with attributes A =
{A1 ; : : : ; Am } appearing in the relations R, we rst ompute the
power-set A∗ onsisting of all subsets of A. Then, the sele tion
operator onstru ts the Ymal queries' onstraints with respe t
to A∗ .
Denition 1: Given a query Q, its attribute power-set A∗
and the result set Res(Q), the sele tion operator outputs, for
ea h set of attributes of the form {Ai ; : : : ; Aj } in A∗ , a sele tion
ondition of the form:
Ai = aix AND : : : AND Aj = ajy ,
Fig. 2. Query results and Ymal re ommendations.
where aix ∈ dom(Ai ), ajy ∈ dom(Aj ) and the set of values
{aix ; : : : ; ajy } is the most frequently appeared value set in MySQL. Implementation on top of an existing database system
Res(Q) for {Ai ; : : : ; Aj }. has a number of advantages, su h as portability and ease of
For ea h sele tion ondition that is returned by the sele tion implementation. Our system an be a essed via a simple web
operator, a dierent Ymal query is onstru ted. browser using an intuitive GUI.
In the latter ase, sele tion onditions are generated taking We demonstrate our method using a movies database. Users
into onsideration not only the query results but also statisti s an submit their queries via SQL or by employing available
maintained for the database ontent. In parti ular, for ea h input forms. After exe uting the query, users are presented with
attribute set in A∗ , we lo ate the relative value set V that the results of the query and also a set of Ymal results (Fig. 2).
appears frequently in both Res(Q) and D using the formula: An explanation is also provided along with ea h Ymal result,
freqRes(Q) (V) where f req Res(Q) (V) denotes the number of o - i.e. how this spe i re ommendation is related to the original
freqD (V) query result. At rst, we present one result tuple of ea h YMAL
urren es of V in Res(Q) and f req D (V) denotes the number of
o urren es of V in D. query. If the user li ks on this, more tuples from the result of
this Ymal query appear as well.
Ymal Query Generator. In this step, we onstru t the Ymal
queries. To perform this operation, we employ the outputs of Referen es

the previous steps. This omponent is also responsible for iden- [1℄ G. Chatzopoulou, M. Eirinaki, and N. Polyzotis. Query re -
tifying the join onditions for ea h generated query. Ymal(Q) ommendations for intera tive database exploration. In SSDBM,
onsists of the union of results of the produ ed Ymal queries. 2009.
[2℄ G. Koutrika, B. Ber ovitz, and H. Gar ia-Molina. Flexre s: Ex-
III. Demonstration pressing and ombining exible re ommendations. In SIGMOD,
2009.
We have implemented our system for omputing a set of [3℄ K. Stefanidis, M. Drosou, and E. Pitoura. \you may also like"
re ommended results for a given query in Java on top of results in relational databases. In PersDB, 2009.