Handbook of Multimedia for Digital Entertainment and Arts- P2 pps

5 Ranked list of movie recommendations Recommendation Algorithms Pure Collaborative Filtering Our collaborative filtering engine applies the typical neighbourhood-basedalgorithm [8], div

16 G Lekakos et al.

Fig 4 Method selection in MoRe

Fig 5 Ranked list of movie recommendations

Recommendation Algorithms

Pure Collaborative Filtering

Our collaborative filtering engine applies the typical neighbourhood-basedalgorithm [8], divided into three steps: (a) computation of similarities betweenthe target and the remaining of the users, (b) neighborhood development and (c)computation of prediction based on weighted average of the neighbors’ ratings onthe target item

1 Personalized Movie Recommendation 17

For the first step, as formula 1 illustrates, the Pearson correlation coefficient isused

r D

Pi



Xi X 

Yi YrP

where Xi and Yi are the ratings of users X and Y for movie I , while X , Y refer

to the mean values of the available ratings for the users X and Y However, inthe MoRe implementation we used formula 2, given below, which is equivalent toformula 1 but it computes similarities faster since it does not need to compute themean rating values n represents the number of commonly rated movies by users Xand Y

r D

nPi

XiYiPi

X iPi

Y is

nPi

X i2

Pi

X i

2s

nPi

Y i2

Pi

Y i

Note that in the above formulas if either user has evaluated all movies with identicalratings the result is a “divide by zero” error and therefore we decided to ignore userswith such ratings In addition, we devaluate the contribution of neighbors with lessthan 50 commonly rated movies by applying a significance weight of n=50 , where

n is the number of ratings in common [32]

At the neighborhood development step of the collaborative filtering process

we select neighbors with positive correlation to the target user In order to increasethe accuracy of the recommendations, prediction for a movie is produced only if theneighbourhood consists of at least 5 neighbors

To compute an arithmetic prediction for a movie, the weighted average of allneighbors’ ratings is computed using formula 3

Ki D NK C

P

J 2Neighbours

.J i  NJ /rKJP

J

where Kiis the prediction for movie i , NK is the average mean of target user’s ratings,

Jiis the rating of neighbour J for the movie i , NJ is the average mean of neighbour

J0s ratings and rKJ is the Pearson correlation measure for the target user and herneighbor J

Pure Content-Based Filtering

In the content-based prediction we consider as features all movie contributors (cast,directors, writers, and producers), the genre, and the plot words Features that appear

18 G Lekakos et al.

in only one movie are ignored Each movie is represented by a vector, the length ofwhich is equal to the number of non-unique features of all available movies Theelements of the vector state the existence or non-existence (Boolean) of a specificfeature in the description of the movie

To calculate the similarity of two movies, we use the cosine similarity measurecomputed in formula 4 ai and bi are the values of the i -th elements of vectorsE

aibirPi

ai2rPi

bi2

(4)

The algorithm we use to produce recommendations is an extension of the top-Nitem-based algorithm that is described by Karypis in [33] Since the movie set doesnot change dynamically when the system is online, the similarities between all pairs

of movies in the dataset are pre-computed off-line and for each movie the k-mostsimilar movies are recorded, along with their corresponding similarity values When

a user that has rated positively (i.e four or five) a set U of movies, asks for mendations, a set C of candidate movies for recommendation is created as the union

recom-of the k-most similar movies for each movie j 2 U , excluding movies already in U The next step is to compute the similarity of each movie c 2 C to the set U as thesum of the similarities between c 2 C and all movies j 2 U Finally, the movies in Care sorted with respect to that similarity Figure6graphically represents the content-based prediction process

Note that typically content-based recommendation is based upon the similaritiesbetween item’ features and user’s profile consisting of preferences on items’ fea-tures Instead, Karypis computes similarities between items upon all users’ ratingscompletely ignoring item features This approach is also known as item-to-item cor-relation and is regarded as content-based retrieval We extend Karypis’ algorithm byutilizing the movies’ features rather than the user’ ratings to find the most similarmovies to the ones that the user has rated positively in the past and therefore wepreserve the term content-based filtering

Since we are interested in numerical ratings in order to combine content-basedand collaborative filtering predictions, we extend Karypis’ algorithm (which is de-

signed for binary ratings) as follows Let MaxSim and MinSim be the maximum and

minimum similarities for each movie in c 2 C to U and S i mi the similarity of amovie Mi to the set U The numerical prediction Prifor the movie is computed byformula 5

Pri D .S i mi MinSim/  4

.MaxSim  MinSim/ C 1 (5)Formula 5 above normalizes similarities from ŒMaxSim; MinSim to Œ1; 5, which

is the rating scale used in collaborative filtering For example, if S i miD 0:8,

MinSim D 0:1, and MaxSim D 0:9 then Pri D 4:5 Note that the formula applies forany similarity value (above or below one)

Due to the fact that movie similarities are computed offline, we are able toproduce content-based recommendations much faster than collaborative filtering

1 Personalized Movie Recommendation 19

Fig 6 Content-based filtering prediction process

recommendations Moreover, in contrast to collaborative filtering, content-basedpredictions can always be produced for the specific dataset

In addition, we implemented content-based filtering using the Na¨ıve Bayes rithm Each of the five numerical ratings is considered as a class label and prediction

algo-u for an item is compalgo-uted algo-using formalgo-ula 6:

, P.uj/ is the

proba-bility that any item can be rated by the user with uj (computed by the available user

ratings), m is the number of terms used in the description of the items and P.aijuj/

is the probability to find in the item’s description the term aiwhen it has been rated

with uj The probability P.aijuj/ is computed by formula 7

Paijuj

D niC 1

20 G Lekakos et al.

where n is the total number of occurrences of all terms that are used for the

descrip-tion of the items and have been rated with uj, nj is the frequency of appearance

of the term aiin the n terms andjVocabularyj is the number of unique terms

ap-pearing in all items that have been rated by the user The Na¨ıve Bayes algorithm hasbeen successfully used in the book recommendation domain [18]

Hybrid Recommendation Methods

The proposed hybrid recommendation method is implemented in two variations

The first one, called substitute, aims to utilize collaborative filtering as the main

prediction method and switch to content-based when collaborative filtering tions cannot be made The use of collaborative filtering as the primary method isbased on the superiority of collaborative filtering in multiple application fields, aswell as in the movie domain [29,30] Content-based predictions are triggered whenthe neighborhood size of the target user consists of less than 5 users

predic-This approach is expected to increase both the prediction accuracy as well as theprediction coverage Indeed, the collaborative filtering algorithm described aboverequires at least five neighbors for the target user in order to make a prediction.This requirement increases the accuracy of the collaborative filtering method itself(compared to the typical collaborative filtering algorithm) but leads to a predictionfailure when it is not met For these items (for which prediction cannot be made)content-based prediction is always feasible and therefore the overall accuracy ofthe substitute hybrid algorithm is expected to improve compared to collaborativefiltering as well as content-based filtering Although this approach is also expected

to improve prediction coverage, the time required to make predictions may increasedue to the additional steps required by the content-based algorithm However, thisdelay may be practically insignificant since the time needed to make content-basedrecommendations is significantly shorter than the time to produce recommendationswith collaborative filtering

The second variation of the proposed hybrid approach, called switching, is based

on the number of available ratings for the target user as the switching criterion laborative filtering prediction is negatively affected when few ratings are availablefor the target user In contrast, content-based method deal with this problem moreeffectively since predictions can be produced even upon few ratings The switchinghybrid uses collaborative filtering as the main recommendation method and trig-gers a content-based prediction when the number of available ratings falls below afixed threshold This threshold value can be experimentally determined and for thespecific dataset has been set to 40 ratings

Col-In terms of prediction coverage the switching hybrid is not expected to differsignificantly from the collaborative filtering prediction, since content-based filter-ing may be applied even if a collaborative filtering prediction can be produced, incontrast to the substitute hybrid which triggers content-based prediction upon the

1 Personalized Movie Recommendation 21

“failure” of collaborative filtering to make predictions Although the two variationsabove follow the exactly the same approach having collaborative filtering as theirmain recommendation method, they differ in the switching criterion

Experimental Evaluation

The objective of the experimental evaluation is to compare the two versions ofthe hybrid algorithm against each other as well as against the base algorithms(collaborative and content-based filtering) The comparison is performed in terms

of predictive accuracy, coverage and actual time required for real-time tions Moreover, since pure collaborative filtering, implemented in MoRe, adopts aneighborhood-size threshold (5 neighbors) we will examine its performance againstthe typical collaborative filtering method without the neighborhood size restriction

predic-We will also demonstrate that the number of features used to describe the moviesplays an important role in the prediction accuracy of the content-based algorithm.The evaluation measures utilized for estimating prediction accuracy is the MeanAbsolute Error (MAE) The Mean Absolute Error [2] is a suitable measure of pre-cision for systems that use numerical user ratings and numerical predictions Ifr1; : : : ; rnare the real values of user in the test, p1; : : : ; pnare the predicted valuesfor the same ratings and E D "1; : : : ; "nD fp1 r1; : : : ; pn rng are the errors,then the Mean Absolute Error is computed by formula 8

MAEDˇˇEˇˇD

nPiD0j"ij

In the experimental process the original dataset is separated in two subsets domly selected: a training set containing the 80% of ratings of each available userand a test set including the remaining 20% of the ratings Furthermore, availableuser ratings have been split in the two subsets The ratings that belong to the test setare ignored by the system and we try to produce predictions for them using only theremaining ratings of the training set To compare the MAE values of the differentrecommendation methods and to verify that the differences are statistically signif-icant we apply non-parametric Wilcoxon rank test, in the 99% confidence space(since normality requirement or parametric test is not met)

ran-The MAE for the pure collaborative filtering method is 0.7597 and the coverage98.34% The MAE value for collaborative filtering method (without the neighbor-hood size restriction) is 0.7654 and the respective coverage 99.2% The p-value

of the Wilcoxon test (p D 0:0002) indicates a statistically significant differencesuggesting that the restriction to produce prediction for a movie only if the neigh-bourhood consists of at least 5 neighbours lead to more accurate predictions, butscarifies a portion of coverage

22 G Lekakos et al Table 1 Number of features

and prediction accuracy Case Threshold(movies) MAE Numberof features

signif-is increased to three, five, ten and fifteen movies, as shown in Table1

Comparing cases 1 and 2 above we notice no significant differences, while thedifference between 2 and 3, 4, 5 (p D 0:0000 for all cases) cases are statisticallysignificant

Thus, we may conclude that the number of features that are used to represent themovies is an important factor of the accuracy of the recommendations and, morespecifically, the more features are used, the more accurate the recommendationsare Note that Na¨ıve Bayes algorithm performed poorly in terms of accuracy withMAE D 1:2434 We improved its performance when considered ratings above 3 aspositive ratings and below 3 as negative (MAE D 1:118) However, this error is stillsignificantly higher than the previous implementation and therefore we exclude itfrom the development of the hybrid approaches

Substitute hybrid recommendation method was designed to achieve 100% age The MAE of the method was calculated to be 0.7501, which is a statistically im-portant improvement of the accuracy of pure collaborative filtering (p < 0:00001).The coverage of the switching hybrid recommendation method is 98.8%, whilethe MAE is 0.7702, which is a statistically different in relevance to substitute hy-brid and pure collaborative filtering methods (p D 0:000) This method producesrecommendations of less accuracy than both pure collaborative filtering and sub-stitute hybrid, has greater coverage than the first and lower that the latter method,but it produces recommendations in reduced time than both methods above Eventhough recommendation methods are usually evaluated in terms of accuracy andcoverage, the reduction of execution time might be considered more important for

cover-a recommender system designer, in pcover-articulcover-ar in cover-a system with cover-a lcover-arge number ofusers and/or items

Table2depicts the MAE values, coverage and time required for real-time diction (on a Pentium machine running at 3.2 GHz with 1 GB RAM) for all fourrecommendation methods

Note that the most demanding algorithm in terms of resources for real-time diction is collaborative filtering If similarities are computed between the target and

pre-1 Personalized Movie Recommendation 23 Table 2 MAE, coverage,

and prediction time for the

run time prediction Pure Collaborative

Filtering

0.7597 98.34% 14 sec Pure Content-based

tions

Recommenda-0.9253 100% 3 sec

Substitute hybrid recommendation method

0.7501 100% 16 sec

Switching hybrid recommendation method

0.7702 98.8% 10 sec

the remaining users at prediction time then its complexity is O nm/ for n usersand m items This may be reduced at O m/ if similarities for all pairs or users arepre-computed with an off-line cost O

n2m However, such a pre-computation stepaffects one of the most important characteristics of collaborative filtering, which isits ability to incorporate the most up-to-date ratings in the prediction process Indomains where rapid changes in user interests are not likely to occur the off-linecomputation step may be a worthwhile alternative

Conclusions and Future Research

The above empirical results provide useful insights concerning collaborative andcontent-based filtering as well as their combination under the substitute and switch-ing hybridization mechanisms

Collaborative filtering remains one of the most accurate recommendation ods but for very large datasets the scalability problem may be considerable and asimilarities pre-computation phase may reduce the run-time prediction cost Thesize of target user’s neighbourhood does affect the accuracy of recommendations.Setting the minimum number of neighbors to 5 improves prediction accuracy but at

meth-a smmeth-all cost in covermeth-age

Content-based recommendations are significantly less accurate than tive filtering, but are produced much faster In the movie recommendation domain,the accuracy depends on the number of features that are used to describe the movies.The more features there are, the more accurate the recommendations

collabora-Substitute hybrid recommendation method improves the performance of orative filtering in both terms of accuracy and coverage Although the difference

collab-in coverage with collaborative filtercollab-ing on the specific dataset and with specificconditions (user rated at least 20 movies, zero weight threshold value) is rather in-significant, it has been reported that this is not always the case, in particular whenincreasing the weight threshold value [32] On the other hand, the switching hybrid

Another issue that could be subject for future research is the way of the mendations presented to the users, the layout of the graphical user interface andhow this influences the user ratings Although there exist some studies on these is-sues (e.g [34]), it is a fact that the focus in recommender system research is on thealgorithms that are used in the recommendation techniques.

recom-References

1 D Goldberg, D Nichols, B.M Oki, and D Terry, “Using Collaborative Filtering to Weave an Information Tapestry,” Communications of the ACM Vol 35, No 12, December, 1992, p.p 61-70.

2 U Shardanand, and P Maes, “Social Information Filtering: Algorithms for Automating “Word

of Mouth”,” Proceedings of the ACM CH’95 Conference on Human Factors in Computing Systems, Denver, Colorado, 1995, p.p 210-217.

3 B N Miller, I Albert, S K Lam, J Konstan, and J Riedl, “MovieLens Unplugged: Experiences with an Occasionally Connected Recommender System,” Proceedings of the In- ternational Conference on Intelligent User Interfaces, 2003.

4 W Hill, L Stead, M Rosenstein, and G Furnas, “Recommending and Evaluating Choices

in a Virtual Community of Use,” Proceedings of the ACM Conference on Human Factors in Computing Systems, 1995, p.p 174-201.

5 Z Yu, and X Zhou, “TV3P: An Adaptive Assistant for Personalized TV,” IEEE Transactions

on Consumer Electronics, Vol 50, No 1, 2004, p.p 393-399.

6 D O’Sullivan, B Smyth, D C Wilson, K McDonald, and A Smeaton, “Improving the Quality of the Personalized Electronic Program Guide,” User Modeling and User Adapted Interaction;Vol 14, No 1, 2004, p.p 5-36.

7 S Gutta, K Kuparati, K Lee, J Martino, D Schaffer, and J Zimmerman, “TV Content Recommender System,” Proceedings of the Seventeenth National Conference on Artificial Intelligence, Austin, Texas, 2000, p.p 1121-1122.

8 P Resnick, N Iacovou, M Suchak, P Bergstrom, and J Riedl, “GroupLens: An Open chitecture for Collaborative Filtering of NetNews,” Proceedings of the ACM Conference on Computer Supported Cooperative Work, 1994, p.p 175-186.

Ar-1 Personalized Movie Recommendation 25

9 J Konstan, B Miller, D Maltz, J Herlocker, L Gordon, and J Riedl, “GroupLens: Applying Collaborative Filtering to Usenet News,” Communications of the ACM, Vol 40, No 3, 1997, p.p 77-87.

10 G Linden, B Smith, and J York, “Amazon.com Recommendations: Item-to-Item tive Filtering,” IEEE Internet Computing, Vol 7, No 1, January-February, 2003, p.p 76-80.

Collabora-11 G Lekakos, and G M Giaglis, “A Lifestyle-based Approach for Delivering Personalized Advertisements in Digital Interactive Television,” Journal Of Computer Mediated Communi- cation, Vol 9, No 2, 2004.

12 B Smyth, and P Cotter, “A Personalized Television Listings Service,” Communications of the ACM;Vol.43, No 8, 2000, p.p 107-111.

13 G Lekakos, and G Giaglis, “Improving the Prediction Accuracy of Recommendation Algorithms: Approaches Anchored on Human Factors,” Interacting with Computers, Vol 18,

16 J Herlocker, J Konstan, and J Riedl, “An Empirical Analysis of Design Choices in Neighborhood-Base Collaborative Filtering Algorithms,” Information Retrieval, Vol 5, No.

19 M Balabanovic, and Y Shoham, “Fab: Content-based Collaborative Recommendation,” munications of the ACM, Vol 40, No 3, 1997, p.p 66-72.

Com-20 M Pazzani, and D Billsus, “Learning and Revising User Profiles: The identification of esting Web sites,” Machine Learning, Vol 27, No 3, 1997, p.p 313-331.

inter-21 M Balabanovic, “An Adaptive Web Page Recommendation Service,” Proceedings of the ACM First International Conference on Autonomous Agents, Marina del Ray, California, 1997, p.p 378-385.

22 M Pazzani, and D Billsus, “Content-based Recommendation Systems,” The Adaptive Web,

25 M Claypool, A Gokhale, T Miranda, P Murnikov, D Netes, and M Sartin,

“Combining Content-Based and Collaborative Filters in an Online Newspaper,” ings of the ACM SIGIR Workshop on Recommender Systems, Berkeley, CA, 1999, http://www.csee.umbc.edu/ ian/sigir99-rec/.

Proceed-26 I Schwab, W Pohl, and I Koychev, “Learning to Recommend from Positive Evidence,” ceedings of the Intelligent User Interfaces, New Orleans, LA, 2000, p.p 241-247.

Pro-27 M Pazzani, “A Framework for Collaborative, Content-Based and Demographic Filtering,” Artificial Intelligence Review, Vol 13, No 5-6, December, 1999, p.p 393-408.

28 R Burke, “Hybrid Web Recommender Systems,” The Adaptive Web, 2007, p.p 377-408.

29 C Basu, H Hirsh, and W Cohen, “Recommendation as Classification: Using Social and Content-based Information in Recommendation,” Proceedings of the Fifteenth National Con- ference on Artificial Intelligence, Madison, WI, 1998, p.p 714-720.

30 J Alspector, A Koicz, and N Karunanithi, “Feature-based and Clique-based User Models for Movie Selection: A Comparative study,” User Modeling and User Adapted Interaction, Vol 7,

no 4, September, 1997, p.p 297-304.

26 G Lekakos et al.

31 A Rashid, I Albert, D Cosley, S Lam, McNee S., J Konstan, and J Riedl, “Getting to Know You: Learning New User Preferences in Recommender Systems,” Proceedings of International Conference on Intelligent User Interfaces, 2002.

32 J Herlocker, J Konstan, A Borchers, and J Riedl, “An Algorithmic Framework for forming Collaborative Filtering,” Proceedings of the Twenty-second International Conference

Per-on Research and Development in InformatiPer-on Retrieval (SIGIR ’99), New York, 1999, p.p 230-237.

33 G Karypis, “Evaluation of Item-Based Top-N Recommendation Algorithms,” Proceedings the Tenth International Conference on Information and Knowledge Management, 2001, p.p 247-254.

34 D Cosle, S Lam, I K Albert, J., and J Riedl, “Is Seeing Believing? How Recommender tems Influence Users’ Opinions,” Proceedings of the SIGCHI Conference on Human Factors

Sys-in ComputSys-ing Systems, Fort Lauderdale, FL, 2003, p.p 585-592.

Chapter 2

Cross-category Recommendation

for Multimedia Content

Naoki Kamimaeda, Tomohiro Tsunoda, and Masaaki Hoshino

of such problems, there is a need for personalization technologies By using suchtechnologies, we can easily find favorite content and avoid storing unnecessary con-tent, because these technologies can select content that interests the user among alarge variety of content

Recommendation services are one of the most popular applications that arebased on personalization technologies Most of these services provide recommenda-tions for individual categories By applying recommendation technologies to severaldifferent categories, user experience can be improved By using user preferences in-volving several categories, the system can figure out more profound nature of user’staste and user’s view point to select content Moreover, it becomes easier to findsimilar content from other categories In this article, this kind of recommendation isreferred to as “cross-category recommendation.”

The purpose of this article is to introduce cross-category recommendation nologies for multimedia content First, in order to understand how to realize therecommendation function, multimedia content recommendation technologies andcross-category recommendation technologies are outlined Second, practical ap-plications and services using these technologies are described Finally, difficul-ties involving cross-category recommendation for multimedia content and futureprospects are mentioned as the conclusion

tech-N Kamimaeda ( ), T Tsunoda, and M Hoshino

Sec 5, Intelligence Application Development Dept., Common Technology Division, Technology Development Group, Corporate R&D, Sony Corporation, Tokyo, Japan

e-mail: Naoki.Kamimaeda@jp.sony.com; tsunoda@sue.sony.co.jp; samba@sue.sony.co.jp

B Furht (ed.), Handbook of Multimedia for Digital Entertainment and Arts,

DOI 10.1007/978-0-387-89024-1 2, c  Springer Science+Business Media, LLC 2009

27

28 N Kamimaeda et al.Technological Overview

Overview

The technological overview is described in two parts: multimedia content endation technologies and cross-category recommendation technologies Therelationship between these technologies is shown in Figure1

recomm-Multimedia recommendation technologies involve basic technologies that can beused to realize recommendation functions for each category Cross-category rec-ommendation technologies involve technologies to realize cross-recommendationamong categories based on multimedia recommendation technologies These twotechnologies have been explained in the following sections

Multimedia Content Recommendation

In this section, an overview of recommendation technologies for multimedia tent is described There are two types of such technologies: collaborative filtering(CF) and content-based filtering (CBF) First, basic technologies about CF are de-scribed Second, we explain CBF technologies in detail, because in this article, wemainly explain cross-category recommendation technologies using CBF technolo-gies After that, typical cases of multimedia content recommendation systems arementioned Finally, how to realize cross-category recommendation based on CBFtechnologies is described

con-Fig 1 Two types of recommendation technologies

2 Cross-category Recommendation for Multimedia Content 29

Basic Technologies Involving CF

Collaborative filtering methods can be categorized into the following two types.One type of CF starts by finding a set of customers whose purchases and rated itemsoverlap the user’s purchases and rated items [2] The algorithm aggregates itemsfrom such similar customers, eliminates items the user has already purchased orrated, and recommends the remaining items to the user This is called user-based

CF Cluster models are also a type of user-based approach

The other type of CF focuses on finding similar items, and not similar customers.For each of the user’s purchased and rated items, the algorithm attempts to find sim-ilar items It then aggregates the similar items and recommends them This is calleditem-based CF Two popular versions of this algorithm are search-based methodsand item-to-item collaborative filtering [3]

Both CF methods cannot often work well with completely new items, items withless reusability such as TV programs, high merchandise turnover rate items, and so

on As a simple example of conventional CF, a problem in TV program dation can be encountered as follows

recommen-1 Tom watched TV programs named X, Y, and Z

2 Mike watched TV programs named X and Y but did not watch Z

3 The system recommends program Z to Mike since Tom and Mike have watchedthe same programs X and Y, but Mike has never watched program Z before

4 However, program Z has already been broadcast and Mike cannot watch program

Z now

Although CF methods have this type of problem, CF can be easily applied tocross-category recommendation, because CF is independent of the type of item,but it depends on which items are purchased or rated together Moreover, tech-nologies using community trends like CF are very important for cross-categoryrecommendation

Lately, several community-based recommendation services have emerged.Last.fm [4], MusicStrands (Mystrands) [5], and Soundflavor [6] are examples

of community-based music recommendation services These sites obtain the ing logs or playlist data of community members; these song playlists are sharedwith other community members and are also used to recommend music

listen-Basic Technologies Involving CBF

Key Elements of a Content Recommendation System Using CBF

A content recommendation system using CBF technologies has four key elements,

as shown in Figure2: content profiling, context learning, user preference learning,and matching

In content profiling, the machine should understand what the content is in order

to recommend it For example, jazz music has acoustic instrumentation and makesfor very relaxed listening Understanding the content seems like an oversimplifica-

30 N Kamimaeda et al.

Fig 2 Four key elements of a CBF-based content recommendation system

tion, but a machine should manage all the necessary information that represents thecontent The next element is context learning Understanding the user’s context isalso important for recommending content The user’s interest is influenced by whereshe/he is, the time of the day, what type of situation she/he is in, or how she/he isfeeling For example, if the user is sitting in a caf´e near a tropical seashore, she/hemay prefer to listen to Latin music with a tropical cocktail in his/her hand Alter-natively, the user may prefer to listen to a wide range of music—classic to punkrock music—in the morning The third element is learning the users’ preferences.Learning and understanding the user’s taste or preference is important to provide ex-cellent recommendation in order to achieve better user satisfaction If a user alwayslistens to songs sung by female vocalists, she/he may prefer vocal to instrumentalmusic The last element is matching Matching methods are used for recommend-ing or searching relevant content This key element measures the relevancy betweenthe three abovementioned entities, such as that between user preference and contentprofile and the similarity between content

In this chapter, these four key elements are discussed in detail; however, let usbriefly introduce other factors such as association discovery, trend discovery (TD),and community-based recommendation TD is useful from the viewpoint of pro-viding recommendations because users often may wish to check the latest populartrends For example, the TD system extracts trends from the World Wide Web(WWW) by employing a text mining technique comprising the following steps: (1)identifying frequent phrases, (2) generating histories or phrases, and (3) seekingtemporal patterns that match a specific trend [7] One research group has focused ondetecting the sentimental information associated with retail products by employingnatural language processing [8]