674 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011

Transcription

1 674 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011 An Automatic Recommendation Scheme of TV Program Contents for (IP)TV Personalization Eunhui Kim, Shinjee Pyo, Eunkyung Park, and Munchurl Kim, Member, IEEE Abstract Due to the rapid increase of contents available under the convergence of broadcasting and Internet, efficient access to personally preferred contents has become an important issue. In this paper, an automatic recommendation scheme based on collaborative filtering is presented for intelligent personalization of (IP)TV services. The proposed scheme does not require TV viewers (users) to make explicit ratings on their watched TV program contents. Instead, it implicitly infers the users interests on the watched TV program contents. For the recommendation of user preferred TV program contents, our proposed recommendation scheme first clusters TV users into similar groups based on their preferences on the content genres from the user s watching history of TV program contents. For the personalized recommendation of TV program contents to an active user, a candidate set of preferred TV program contents is obtained via collaborative filtering for the group to which the active user belongs. The candidate TV programs for recommendation are then ranked by a proposed novel ranking model. Finally, a set of topranked TV program contents is recommended to the active user. The experimental results show that the proposed TV program recommendation scheme yields about 77% of average precision accuracy and value of (Average Normalized Modified Retrieval Rank) with top five recommendations for 1,509 people. Index Terms Collaborative filtering, content based filtering, TV personalization, TV program recommendation. I. INTRODUCTION DUE to the convergence of broadcasting and internet, the number of TV program contents available at user sides is rapidly increasing and the accessibility to the TV program contents becomes an important issue in TV watching environments of traditional TV, IPTV or TV portals services. Therefore, it is important for users (TV viewers) to easily find and access their preferred ones from TV program contents available to their terminals. There are two approaches to the provision of user s preferred TV program contents; content searching and Manuscript received August 31, 2010; revised February 11, 2011; accepted June 27, Date of publication August 08, 2011; date of current version August 24, This work was supported by the R&D program of MKE/IITA (A , Development of Open-IPTV Technologies for Wired and Wireless Networks). E. Kim and E. Park are with the Dept. of Electrical Engineering at Korea Advanced Institute of Science and Technology (KAIST), Daejeon , Korea ( lins77@kaist.ac.kr; epark@kaist.ac.kr). S. Pyo is with the Dept. of Information and Communications Engineering at KAIST, Daejeon , Korea ( sjpyo@kaist.ac.kr). M. Kim is with the Dept. of Electrical Engineering at KAIST, Daejeon , Korea ( mkim@ee.kaist.ac.kr). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /TBC recommendation. For content searching, TV users usually input the query words into search engines via graphical user interface (GUI), and get the search results from which they finally select their preferred ones. The disadvantage of content searching is that the users do not even know the keywords to search what they want. For content recommendation, it can be possible to recommend to the users their preferred TV program contents. Content recommendation based approach can greatly alleviate user s burden to access their preferred TV program contents. That is, the number of interactions to TV program GUI can be reduced which is often the case in the traditional (IP)TV environments. We use the term (IP)TV as IPTV and conventional TV in this paper. Collaborative filtering (CF) has often been used to recommend goods for e-commerce in Internet. The main idea of CF techniques is based on item preferences of similar users [1] [3]. However, the CF has the following characteristics: (1) it is often designed to recommend to the active users (the target users for recommendation) the items which have not been purchased (consumed) before. This is not always appropriate for TV viewers because they are often likely to watch the same TV program series such as drama, series, and news etc which are repeatedly broadcast; (2) high computational complexity is caused to deal with many users and items. The content based filtering (CBF) has been used in information filtering [1] [3]. The CBF usually recommends items based on the previously evaluated description by the active users. The weak points of CBF are as follows: (1) its recommendation is restricted to the items that the active user has rated or consumed. In other words, it is over-specialized recommendation which heavily depends on user s consumption history on items; (2) it also requires heavy computational complexity for reliable recommendation in analyzing various characteristics of the items that users have consumed [6]. In this paper, we present a personalized automatic recommendation scheme of TV program contents based on CF. The proposed recommendation scheme consists of three parts user profile reasoning, user clustering and recommendation of TV program contents. Unlike the traditional CF-based item recommendation that requires explicit ratings on the purchased items by users, the proposed recommendation scheme implicitly learns the user s interest on the TV program contents and genres, which does not require user s explicit ratings on their watched TV program contents, thus making it more practical in real TV watching environments. For user clustering, similar user groups are clustered based on the feature vectors of TV program genres computed from the usage history of the watched TV program contents by users. Two methods of user /$ IEEE

2 KIM et al.: AUTOMATIC RECOMMENDATION SCHEME OF TV PROGRAM CONTENTS FOR (IP)TV PERSONALIZATION 675 clustering are compared in this paper: demographic clustering and -means clustering. Finally, CF -based recommendation is performed with a novel ranking model which extends the Best Match (BM) model [7] [9] to rank the candidate TV program contents for recommendation. The proposed rank model is designed to make easy access to preferred TV program contents. For the recommendation of popular or newly broadcast TV program contents, the popular TV program contents can be identified via CF from similar user groups. On the other hand, newly broadcast TV program contents are preferably recommended by restricting not recently broadcast TV program contents outside a sliding time window in the history data of watched TV program contents. This paper is organized as follows: Section II reviews the previous related works for recommendation; Section III introduces the overall system architecture of our proposed automatic recommendation scheme for TV program contents and describes the data used for experiments; Section IV describes the components of the proposed recommendation scheme in detail user profile reasoning, similar user clustering, and ranking of candidate TV program contents for recommendation; In Section V, the proposed rank model is explained in detail; The experimental results are presented in Section VI; Finally, Section VII concludes this work. II. RELATED WORKS PTV [4] adopted a hybrid method of CF and CBF to supplement the item ramp-up problem of CF and the user ramp-up problem of CBF [1]. It requires users to provide their preference information on contents while enrolling. Based on this preference information provided by the users, it creates and manages user profiles with explicit ratings by users. However, in general users do not want to offer their personal information or sometimes do not faithfully exhibit their interests on the items with explicit ratings. Pazzani et al. reported that only 15% of people respond to the request for the relevance feedback on their preference [5]. Therefore, requiring users to rate explicitly on items is one of the main reasons that cause rating sparsity problems [3]. Deshpande M. et al. proposed an item based top- recommendation algorithm [11] and J. Wang et al. [12] proposed an extension to relevance model from language model [13], both of which utilize CF with user-profile and item matching for item recommendation. The item based top- recommendation algorithm uses an item-to-item matrix for item recommendation which is computed based on a user-to-item matrix [11] for which the recommendation performance is further improved by extending a language model to a relevance model [12]. In item recommendation of e-commerce, recommender systems tend to suggest new items to the users because they are not likely to repurchase the same items or similar kinds after they have bought them. However, this may not be appropriate in TV environments where TV viewers are expected to watch (consume) the TV program contents (items) that they have been accustomed to watch. In general, TV viewers tend to watch popular TV program contents as their similar taste users do or specific TV program contents according to their individual preferences. So, the previous two models are insufficient in that the more frequently watched TV program contents by an active user are recommended in lower ranks. The proposed rank model in this paper tries to remedy this weak point of the previous rank models. In the CF systems, for a large number of users, the process of grouping similar users and recommending items entails a computational complexity issue [1], [3]. To solve the system overload in clustering similar taste users for a large number of users, G.R. Xue suggests a two-step clustering method by using offline -means clustering and online PCC (Pearson correlation coefficient) clustering for the item ratings [16]. In this method, the -means clustering is performed offline for a large number of users just one time for a given value. Then more similar users are extracted online for active users based on PCC values from their respective clusters to predict the rating values for the unrated items. However, this method needs to know an appropriate cluster number a priori. For user clustering in this paper, we compare two clustering methods: demographic clustering and -means clustering. The demographic clustering is very simple to only use user s demographic information such as genders and ages for clustering. For the -means clustering, we use the feature vectors of user preference values of 8 genres and 47 subgenres for TV program contents. The former one is computationally very simple and can be a solution for the cold-start problem which takes time to learn users, but requires a priori knowledge about the demographic information. On the other hand, the latter one does not require such demographic information but implicitly clusters similar users based on the genre preference from the watched history data of TV program contents by users. For the -means clustering, an appropriate number for can be found by searching a range based on dendrogram of hierarchical clustering [15], [17], [18]. We then determine a value based on the smallest sum of squared errors in this paper. The details of finding a right value are described in Section IV. And as a rank model for ordered recommendation of TV program contents, we propose a novel rank model based on the BM model [7] [9]. The proposed rank model is described in Section V in detail. We can summarize the contribution points of our personalized automatic recommendation scheme for TV program contents as follows: (1) it is more appropriate for TV program recommendation since it makes implicit reasoning for user preference on TV program contents from the watched TV program history data, which does not require users to explicitly rate their watched TV program contents; (2) it takes into account not only the group preferences but also the individual user s preferences on TV program contents for recommendation; and (3) the proposed rank model elaborates collaborative filtering by considering the relative lengths of watching times for TV program contents, not just by simply counting the number of users who have watched them. III. ARCHITECTURE OF THE PROPOSED RECOMMENDATION SYSTEM AND EXPERIMENT DATA A. Proposed Recommendation System Scheme In this paper, it is assumed that TV terminals are connected to the content servers of TV programs via back channels so that

3 676 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011 TABLE I FIELDS OF TV USAGE HISTORY DATABASE Fig. 1. Architecture of the proposed recommendation system for TV program content. the usage (or watching) history of (IP)TV program contents can be collected at the server sides. In IPTV environments, TV program contents are streamed over IP networks and the responsible content providers at head-end sides can collect usage history of TV programs watched by the users via back channels. Fig. 1 shows the architecture of our proposed automatic recommendation system for TV program contents. The automatic recommendation scheme consists of three agents: (1) the user profile reasoning agent computes user preferences on genres and TV programs by analyzing user s watching history of TV program contents. So, this agent collects TV usage history from local repositories of TV terminals for user profile reasoning; (2) the user clustering agent clusters the users (TV viewers) into similar preference user groups; (3) the recommendation agent recommends to each active user a list of his/her preferred TV program contents. Here, an active user means the user who logs into the TV terminal and is ready to receive a recommended TV program list. For recommendation, a list of candidate TV program contents is extracted based on CF and our proposed rank model calculates the respective scores of the candidate TV program contents for ranked recommendation. Then the TV program contents with the top highest scores are presented to the active user in a descending rank order as the result of recommendation. Notice in this paper that the users and items are interchangeably used with the TV viewers and TV program contents, respectively. B. Description of Usage History Data Set for Watched TV Program Contents For the experiments to test the effectiveness of the proposed recommendation scheme, Neilson Korea s TV usage history data set of 2,005 people is used which has been collected on 6 terrestrial TV channels for 6 months from Dec. 1, 2002 to May 31, Table I shows the data fields of the usage history data set for watched TV program contents. The TV program contents in the history data set have 8 main genres which are further divided into 47 subgenres in total. For the data set, the total number of TV program titles amounts to 924 and the total number of subtitles is 1,855. We use 795 TV program contents for training, corresponding to the first 4 months and 629 TV program contents for testing, corresponding to the last 2 months. Notice that the sum (1,424) of the TV program contents for training and testing exceeds the total number (924) of the titles because 500 watched TV program contents are the titles that were repeatedly broadcast. Table I shows the information attached to the broadcast TV program contents. IV. PROPOSED RECOMMENDATION SCHEME A. User Profile Reasoning In this paper, a user is characterized in terms of his/her profile which consists of two preferences on items (TV program contents) and genres. First of all, we remove from the usage history data set all the TV program contents that have not been watched for more than 10% of their respective total lengths. The preference on a TV program content is defined as the relatively watched ratio over the total time length. For the reruns of TV program contents, they are all considered the same title (item). The preference on item by user is defined as where is the number of times being broadcast for an item. And is defined as where and indicate the watched time length for item by user and the total length of an item, respectively. It must be pointed out in (1) that the item preference might be inaccurately computed for inattentively watched TV program contents. The treatment of them is out of scope in this paper. (1) (2)

4 KIM et al.: AUTOMATIC RECOMMENDATION SCHEME OF TV PROGRAM CONTENTS FOR (IP)TV PERSONALIZATION 677 Since the popularity or recency of TV program contents are often diminished with time and the user interest on TV program contents varies over time, it is more appropriate to reflect the recently watched TV program contents for recommendation. Therefore, a time window function is defined as TABLE II SELECTION RESULTS OF CLUSTER NUMBERS, K (3) where is a control parameter for the window size which is set to two-month or four-month length in this paper. The average of user preference on item by user is given by where is the total number of items in the watched TV program list by user. In order to efficiently perform similar user clustering in low dimension, genre preference is used which can reflect the similarity of user s content consumption for TV program contents. Genre preference is computed by accumulating the item preference values for the genre and is then normalized by the total genre preference values for all genres. When the total number of genres is, the genre preference on genre by user is defined as B. User Clustering For computational efficiency and effectiveness of collaborative filtering, TV users are clustered into similar user groups. After clustering, each user has a membership to one of the user groups. Therefore, CF operation for an active user is performed for the user group to which the user belongs, not for the whole users. For similar user grouping, two clustering approaches are compared: demographic clustering based on genders and ages, and -means clustering based on the genre preference as described in (5). For -means clustering, two feature vectors are compared with 8 preferences on the main genres and 47 preferences on the subgenres, respectively. The demographic clustering is computationally very simple but can only be used if the demographic information such as genders and ages is available. The demographic clustering can avoid the cold-start clustering problem that usually takes time while learning the users. In our demographic clustering, there are 26 combinations of different genders and ages. The genders are divided into two classes male and female and the ages are divided into 13 classes, and 66 ages and higher. As an alternative, -means clustering can be used, which does not require the demographic information. The essential prerequisite for -means clustering is to know an appropriate number as the number of clusters. In order to find a right (4) (5) and reveal the characteristics of cluster distributions, we take a two-step approach: an unsupervised hierarchical clustering is first run to construct a dendrogram for which a range of values is found by cutting its branches at the large jumps in a distance criterion [14], [15], [17], [18]; the final value is then determined in the range by repeatedly performing -means clustering. To determine the final value, -means clustering is repeated times for which the centroids of clusters are randomly initialized each time. When the clustering yields the same clustering results times for a given value, the clustering results become the final clusters with the value. When any does not result in the same clustering results less than times, the value that results in the same clustering results the largest times is selected as the final value. In this paper, and are set to 1,000 and 5, respectively. Table II shows ranges and finally selected values for the features vectors of 47 subgenres and 8 main genres for the watched TV program contents by the users who have watched at least 33% of the average number of watched TV program contents during the training period. For this experiment, the open-source code Cluster 3.0 was used in [14]. The K-means clustering, which is the most time consuming task in our scheme, takes less than one minute for on a PC with Intel Core 2 Quad CPU 2.4 GHz and 2 GB memory. C. Recommendation of TV Program Contents In order to recommend the preferred TV program contents to an active user, the recommendation process consists of three steps: extracting similar preference users from the clusters (similar user groups) to which the active user belongs; selecting candidate items for recommendation; and ranking the candidate items. Especially the rank model will be explained in Section V in details. 1) Selecting Similar Preference Users of an Active User in a Group: In Section IV-B, clustering the similar preference users is done offline. For recommendation, more relevant users are further extracted to construct a set of candidate TV program contents based on CF for the user group to which an active user belongs. By doing so, the computation complexity is lowered by reducing the number of all users to the number of similar peer users in the similar user group to which the active user belong. Based on the proximity measure, the most peer users with similar preference are extracted for the active user. For the proximity measure, the normalized correlation is computed by subtracting the average preference value from all the preference values [10]. On the basis of the consumed (watched) item (TV program contents) list of an active user, the similarity between and each peer user is measured as the proximity between

5 678 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011 the normalized preference values on items for and in the similar preference groups. The similarity is defined as (6) where is an item belonging to. represents the active user s profile and indicates the preference value on item consumed by user as in (1), and is the averaged item preference value of user as in (4). The users with are only regarded as relevant users to the active user. Then CF is performed for the item lists between the active user and each of the relevant users. Since the number of similar preference users affects precision performance, we need to find an optimal number of peer users based on the average precision accuracy, which is explained in Section VI. 2) Filtering Candidate Items With EPG Information: After selecting the relevant users for an active user, their preference items become the candidate items for recommendation. But some items may not be available in TV channels due to the termination of broadcasting for the TV program contents. In case of linear TV broadcasting services, Electronic Program Guide (EPG) information can be used to filter out the candidate TV program contents which are not available. 3) Ranking Items: After a set of candidate TV program contents for recommendation is determined, they are ordered by a rank model. Finally, the recommended TV program contents are presented to active users in the descending order of rank scores. The proposed rank model is described in the following section. V. PROPOSED RANK MODEL A. Related Work BM Model Our proposed ranked model extends the Best Match (BM) model [7] [9]. The BM is a ranking function used by retrieval engines to rank matching documents according to their relevance to a given query. The BM model is given by where In (8), is the number of total documents, is the number of documents including a specific term of query, is the number of documents related with a specific topic, and is the number of documents including a specific term of query and is related with the specific topic [8]. In (7), is term frequency in documents and is term frequency in the query. The BM model originates from two Poisson models that the term frequency is independent of relevant and irrelevant documents [9]. Based on this idea, the simple formation is suggested under the following (7) (8) two conditions: one is that the weight is independent of term frequency; and the other is that the weight is linear with term frequency. Each condition is satisfied as and [9]. But, the second condition is not always satisfied as. To remedy this, a scaling factor is added in the numerator, thus resulting in. This is taken into account in the BM11, 15 and 25 models [7]. The BM25 includes an inappropriate condition for TV program recommendation since it gives a high weight on short documents compared to long documents by scope hypothesis [7]. Therefore, the proposed rank model in this paper extends the BM15 model which does incorporate the scope hypothesis into its rank model so that the TV program contents that were broadcast less times are prevented from being higher-ranked. B. Proposed Rank Model An extension to the BM15 is made by taking into account the collaborative filtering concept that accounts for the watching times of users in the rank model for recommendation of TV program contents. Furthermore, we add to the rank model a weight with the correlation between candidate items for recommendation and the items watched by the active user. We score the filtered candidates of TV program contents in a ranked order. The relations between candidate document and query in BM15 are translated into the relations between candidate TV program contents for recommendation and the active user. To make the BM15 be applicable for recommendation of TV program contents, we have the following assumptions: (1) the watched TV program list represents the active user ; (2) is transformed into the relative watching frequencies of both TV program contents of similar preference users and of an active user by applying CF concept, where indicates the watched TV program contents by ; (3) is regarded as the relative watching frequency of by ; (4) the similarity between and by is further taken into account. The matching score between and is defined as our proposed rank model by where and are used to balance the term frequency and the query term frequency in the rank model. The Robertson et al. analyzed the way of weighting in details [7]. and are set to 200 and 0.2 empirically in this paper. In (9), indicates the relative watching frequency which is the ratio of the total number of watching times of both programs and over the total number of watching times of the TV program contents (all s) by the peer users. Therefore, the relative watching frequency is calculated as (9) (10)

6 KIM et al.: AUTOMATIC RECOMMENDATION SCHEME OF TV PROGRAM CONTENTS FOR (IP)TV PERSONALIZATION 679 items. There are two users, and, who have watched two items (TV program contents) and, and the similarity (VCC) value between and is On the contrary, for the two users ( and ) who have watched both and, the VCC value between and is 0.4. So, if we set 0.5 of the VCC value as a threshold for the similarity between items, then the items and are considered being similar, but and are not similar. So, in (9) can improve the rank model by taking into account the relation between the active user and the candidate items for the score calculation. The effect of on precision performance will be shown in Fig. 5 in Section VI. Fig. 2. Illustration for significance on weights w. In (9), indicates the ratio of the total number of watching times of TV program contents over the total number of watching times for all the TV program contents (all s) by, and is given by (11) Eq. (9) can be explained intuitively as follow: is regarded as the peer users preferences on TV program contents in the same user group to which belongs; and is referred to as the active user s preference on TV program contents. The two terms and are in mutually supplemental relation as being multiplied together. In (9),. Two weights and are given as (12) (13) In (12), indicates the total number of broadcast times for all items and is the number of broadcast times of each item. reflects the inverse document frequency with independence assumption between the documents with and without the terms [8]. In this paper, it is assumed that the document for retrieval is and the specific term of query is from active user profile. is added as a weight for the similarity between and which is calculated as vector cosine correlation similarity in (13) for which the and are the feature vectors of user preference on program and, respectively. This weight puts more emphasis on the active user s personal preference on TV program contents, which is not reflected in the original BM [7], [8]. In order to see the effectiveness of in (13), Fig. 2 illustrates an example of similarity measures between two VI. EXPERIMENTAL RESULTS For the usage history of watched TV program contents explained in Section III-B, we use the usage history data of four months for training and the remaining two months for testing. In this experiment section, we measure the performance of our recommendation scheme in terms of both precision/recall and Average Normalized Modified Retrieval Rank (ANMRR) which considers the rank orders in retrieval [19] [21]. A. Performance Measure of Rank Models Precision and Recall: The performance in information retrieval is usually measured in terms of precision and recall [22]. The precision is defined as the ratio of how many watched TV program contents (relevant documents) are contained in the recommendation list (retrieved documents) of TV program contents for an active user. The recall is defined as the ratio of how many recommended TV program contents (retrieved documents) are actually included in the watched TV programs (relevant documents) for the active user. The precision and recall are defined as. (14) (15) where is the number of watched TV program contents in the recommended list of TV program contents and is the number of recommended TV program contents. is the number of recommended TV program contents in the watched list of TV program contents, and is the number of watched TV program contents. For the recommendation of TV program contents, the precision is a more appropriate metric for performance evaluation than the recall since the recall accuracy is increased as the number of recommended TV program contents increases. In this regard, recommending a larger number of TV program contents increases false positives. So, in this paper, we use precision accuracy for performance evaluation. However, if the number of ground truth increases, the precision also becomes higher. So, the performance of the rank models is measured in terms of precision and recall. ANMRR: Compared to precision measure, another performance measure, ANMRR [19] [21], is considered which has been developed to measure the image retrieval performance in

7 680 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011 Fig. 3. Preferences on genres and channels for groups: Demographic Clustering (DM) vs. K-means clustering (KM). (a) Genre preferences of groups. (b) Channel preferences of groups. MPEG-7 [20]. The ANMRR indicates not only how many correct items are recommended but also how highly more relevant items are ranked among the recommended items. For ANMRR, Normalized Modified Retrieval Rank (NMRR) is defined as (16) where is the number of recommended TV program contents that the active user has really watched longer than the average watching times of his/her preferred TV program contents during test period. is the allowable maximum rank and is computed as where is the maximum of [21]. And the in (16) is revised by (17) where is the rank ordered in score values by the proposed rank model in this paper. Finally ANMRR is written as follows: (18) B. Clustered Data Analysis As explained before, two clustering methods are compared between the demographic clustering and -means clustering. A cluster s preference on a specific genre is computed by accumulating the preferences on the specific genre by all users in the same cluster, and then it is normalized by the total number of users in the cluster. Similarly, the normalized preference on a channel can also be computed for each cluster. The preferences and on genre and channel for a cluster are calculated as (19) (20) where is the total number of users in the cluster. and are the total numbers of genres and channels, respectively. Fig. 3 shows the profiles of clusters preferences on genres and channels of TV programs. As shown in Fig. 3, the genre preferences are not significantly distinguished among different groups by demographic clustering (DM). On the other hand, the groups by -means clustering (KM) show somewhat different patterns for genre preferences among them. This is also similarly observed for the channel preferences except the group4 and group5 by DM. Table III shows the average precision performance for different numbers of groups by DM and KM. Although the preferences on genres and channels are better distinguished by KM than DM for different groups, the performance difference of average precision between DM and KM is very slight. In this experiment, the KM turns out to be effective for recommendation

8 KIM et al.: AUTOMATIC RECOMMENDATION SCHEME OF TV PROGRAM CONTENTS FOR (IP)TV PERSONALIZATION 681 TABLE III COMPARISONS OF AVERAGE PRECISION BETWEEN DM AND KM TABLE V PRECISION ACCURACY WITH OUTLIER REMOVALS outlier criteria (33%); refer to Table V. the number of peer user is 5; refer to Fig. 8. TABLE IV AVERAGE PRECISION PERFORMANCE FOR THE NUMBER OF CLUSTERS the number of cluster is 26 by DM; refer to Table IV. the number of peer user is 5; refer to Fig. 8. outlier criteria (33%); refer to Table V. the number of peer user is 5; refer to Fig. 8. Fig. 5. Performance comparison with/without w. Fig. 4. Number of users versus relative watching lengths of TV program contents. although though it does not utilize the demographic information for clustering. Table IV shows the average precision performance on Toprecommendations for different numbers of clusters (groups) by DM. Increasing the number of clusters does not enhance the precision performance because we only use the most similar users to an active user of his/her group (The average precision performance according to the number of peer users will be shown in Fig. 8). C. Exclusion of Nosy Items and Outliers of Users For the experiments, two kinds of outliers are removed to have reliable recommendation: firstly, the TV program contents that were watched less than 10 % of their respective whole lengths are removed as noise; secondly, the users who have watched TV program contents less than a predefined TV watching times are excluded. Fig. 4 shows the number of users versus the relative watching lengths for the two TV program contents Hometown at 6 o clock and Let s marry. For both TV program contents, there are relatively large numbers of users who have watched them less than 10% or more than 95% of the total TV program lengths, respectively. This pattern is similarly observed in other TV program contents. So, we set 10% of the total length of TV program contents as a threshold for outlier removal. Table V shows the average precision performance on different thresholds of outlier removal for the second case. With the exclusion of users who watched the TV program content less than the 33% of the average number of watched TV program contents, we obtain 76.6% of average precision accuracy for the Top-5 recommendation. The threshold values in Table V indicate the ratios of the number of watched TV program contents by each user over the average number of watched TV program contents by all users during the training period of 4 months. The average numbers of watched TV program contents by all users are 124 and 90 during the training period of 4 months and the testing period of 2 months, respectively. D. Effect of on Precision Performance of the Proposed Rank Model Fig. 5 shows the performance comparison in terms of average precision accuracy for the recommended TV program contents with and without in (9). The average precision accuracies with in the proposed rank model are higher than those without it. The average precision in this experiment is measured with 67 active users. E. Performance Comparison Between Proposed Rank Model and Linear Model For performance comparison in precision and ANMRR between the proposed rank model and the linear model [12], 90

9 682 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011 Fig. 6. Performance comparison in precision and recall. Fig. 8. Precision accuracies versus different numbers of similar peer users. (a) Precision accuracies with Top-5 recommendations. (b) Precision accuracies with four clusters. Fig. 7. Performance comparison in ANMRR. users are randomly selected. Figs. 6 and 7 show the performance comparisons of the proposed model and linear model in precision-recall and ANMRR, respectively. The proposed rank model outperforms the linear model in both precision-recall and ANMR. Notice that the smaller the ANMRR value is, the better the recommendation performance is. Ideally, the case of is achieved when the ranked order of the recommended TV program contents is perfectly matched with the order of the watched TV program contents by the active user during the test period. Therefore the recommended TV program contents by the proposed rank model are also better matched in ranked orders than the linear model. The superiority of our proposed rank model comes from the facts that: 1) The proposed rank model defines the weight in (12) such that more frequently broadcast TV program contents are put in lower ranks; 2) For recommendation of TV program contents, the traditional models usually intensify the preference of the peer users but relatively reduce the preference of an active user, which might be appropriate to recommend unpurchased items to active users in e-commerce environments. However, in TV environments, users often tend to watch the TV program contents that they used to watch. Therefore, it is reasonable to take into consideration the preferences of both similar users and an active user for recommendation. The proposed rank model actually considers both; 3) It takes into account the number of watching times for both TV program contents and by peer users in the score calculation for ranking. This more elaborates collaborative filtering. On the other hand, the linear rank model [12] simply counts the number of peer users who have watched both and. F. Performance Analysis for Proposed Recommendation Scheme We investigate the performance of the proposed recommendation in terms of the number of clusters, the number of similar peer users and the number of TV program contents for final recommendation. Fig. 8 shows precision performance for different numbers of similar peer users given Top-5 recommendations and 4 clusters. In Fig. 8(a), the average precision performance slightly decreases as the number of similar peer users increases for different numbers of clusters. This is because a smaller number of similar peer users yields more correlation between the active user and peer users so that the resulting recommendation precision is usually enhanced. When the number of clusters increases, the resulting recommendation precision seldom varies for different numbers of recommended items, because the larger the number of clusters, the more correlate the clustered users are. In Fig. 8(b), the average precision performance of the proposed recommendation scheme becomes lowered as the number of recommended TV program contents increases. Table VI shows the 19 recommended TV program contents by the proposed rank model for the corresponding ground truth items out of 67 for an active user with.as

10 KIM et al.: AUTOMATIC RECOMMENDATION SCHEME OF TV PROGRAM CONTENTS FOR (IP)TV PERSONALIZATION 683 TABLE VI RECOMMENDATION RESULTS AND GROUND TRUTH FOR AN ACTIVE USER WITH ID = # recommendation order, ## preference order of the active user. aforementioned, the more frequently watched TV program contents such as daily news, daily soap opera and weekly regular drama are shown to appear higher-ranked. So this can help active users easily to access their frequently watching TV program contents. On the other hand, the low-ranked items by the proposed rank model are the TV program contents that were not often or never watched by the active user but frequently watched by his/her peer users via the incorporation of collaborative filtering into recommendation. VII. CONCLUSION In this paper, we propose an automatic recommendation scheme of (IP)TV program contents for TV personalization. Unlike the tradition recommendation in document retrieval or e-commerce, the proposed scheme does not require the explicit ratings on watched TV program contents, rather making implicit reasoning for user preference on TV program contents in the usage history data of watched TV program contents. The rank model in the proposed scheme takes into account not only the group preferences but also the active user s preferences on TV program contents. Furthermore, the proposed rank model elaborates collaborative filtering by considering the relative lengths of watching times for TV program contents, not just by simply counting the number of users who have watched them. Our proposed recommendation scheme shows the effectiveness with rich experimental results for a real usage history dataset of watched TV program contents. REFERENCES [1] M. Montaner, B. Lopez, and J. L. DE Larosa, A taxonomy of recommender agents on the Internet, AI Rev., vol. 19, pp , [2] R. Bruke, Hybrid recommender systems: Survey and experiments, User Modeling and User-Adapted Interaction, vol. 12, no. 4, pp , Nov [3] G. Adomavicius and A. Tuzhilin, Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Trans. Knowl. Data Eng., vol. 17, no. 6, pp , Jun [4] P. Cotter and B. Smyth, PTV: Intelligent personalized TV Guides, Amer. Assoc. AI, pp , [5] M. Pazzani and D. Billsus, Learning and revising user profiles: The identification of interesting web sites, Machine Learning, vol. 27, pp , [6] N. Good, J. B. Schafer, J. A. Konstan, A. Borchers, B. Sarwar, J. Herlocker, and J. Riedl, GroupLens research project, combining collaborative filtering with personal agents for better recommendations, Amer. Assoc. AI, [7] S. E. Robertson, S. Walker, M. Beaulieu, M. Gatford, and A. Payne, Okapi at TREC-4, in 4th Text Retrieval Conf. (TREC-4), 1995, pp [8] S. E. Robertson and K. Spark Jones, Relevance weighting of search terms, J. Amer. Soc. Inf. Sci., vol. 27, pp , [9] S. E. Robertson and S. Walker, Some Simple Effective Approximations to the 2-Poisson Model for Probabilistic Weighted Retrieval. New York: Springer-Verlag, 1994, pp [10] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl, Grouplens: An open aarchitecture for collaborative filtering of netnews, in ACM Conf. Comput. Supported Cooperative Work, 1994, pp [11] M. Deshpande and G. Karvpis, Item-based top-n recommendation algorithms, ACM Trans. Inf. Syst., vol. 22, no. 1, pp , Jan [12] J. Wang, J. Powelse, J. Fokker, A. Vreies, and M. Reinders, Personalization on a peer-to-peer television system, Multimedia Tools Appl., vol. 36, no. 1/2, pp , [13] J. Lafferty and C. Zhai, Probabilistic relevance models based on document and query generation, Language Modeling Inf. Retrieval, [14] M. J.L. De Hoon, S. Imoto, J. Nolan, and S. Miyano, Open source clustering software, Bioinfomatics, p. 781, [15] D. P. Vetrov and L. I. Kuncheva, Evaluation of stability of K-means cluster ensembles with respect to random initialization, IEEE Trans. PAMI, vol. 28, no. 11, pp , [16] G. Xue, C. Lin, Q. Yang, W. Xi, H. Zeng, Y. Yu, and Z. Chen, Scalable collaborative filtering using cluster-based smoothing, in ACM SIGIR, Aug. 2005, pp [17] T. Sergios and K. Konstantions, Pattern Recognition, 3rd ed. New York: Academic Press, 2006, pp [18] R. Duda, P. Hart, and D. Stork, Pattern Classification, 2nd ed. New York: Wiley-Interscience, 2001, pp [19] B. S. Manjunath, J.-R. Ohm, V. V. Vasudevan, and A. Yamada, Color and texture descriptors, IEEE Trans. Circuits Syst. Video Technol., vol. 11, no. 6, pp , Jun [20] P. Ndjiki-Nya, J. Restat, T. Meiers, J.-R. Ohm, A. Seyferth, and R. Sniehotta, Subjective evaluation of the MPEG-7 retrieval accuracy measure (ANMRR), in ISO/WG11 MPEG Meeting, Geneva, Switzerland, May 2000, Doc. M6029. [21] W. Ka-Man and P. Lai-Man, MEPG-7 dominant color descriptor based relevance feedback usingmerged palette histogram, in IEEE Int. Conf. Acoust., Speech, Signal Process., May 2004, vol. 3, pp [22] C. D. Manning, P. Raghavan, and H. Schütze, Introduction to Information Retrieval. Cambridge, U.K.: Cambridge Univ. Press, 2008, pp EunHui Kim received the B.E. degree in information and communications engineering from Chungnam National University in 2000 and the M.Sc. degree in information communications engineering from Korea Advanced Institute and Science Technology (KAIST), Daejeon, Korea in She is currently pursuing the Ph.D. degree in Department of Electrical Engineering at KAIST. She worked for Samsung Electronics as an Assistant Engineer of Software team in Visual Display Division during in Suwon, Korea and as an Associate Engineer of Architecture team in Digital Solution Center during in Seoul Korea. Her research interests include personalization in connected TV, data clustering, collaborative filtering, and recommendation modeling with AI for smart TV interaction.

11 684 IEEE TRANSACTIONS ON BROADCASTING, VOL. 57, NO. 3, SEPTEMBER 2011 Shinjee Pyo received the B.E. degree and the M.Sc. degree in information and communications engineering from KAIST, Daejeon, Korea, in 2007 and 2009, respectively. She is currently pursuing the Ph.D. degree in information and communications engineering at KAIST. Her research interests include Personalization in Connected TV, sequential pattern mining for TV personalization and pattern recognition. Eunkyung Park received the B.E. degree in information and communications engineering and the M.Sc. in electrical engineering from KAIST, Deajeon, Korea, in 2009 and 2011, respectively. Now she joins NAVER which is the first and largest search portal in Korea and is working with business and planning for web portal services. Her research interest is statistical learning theory, social networking, and machine learning. Munchurl Kim (M 07) received the B.E. degree in electronics from Kyungpook National University, Korea in 1989, and the M.E. and Ph.D. degrees in electrical and computer engineering from University of Florida, Gainesville, Florida, in 1992 and 1996, respectively. After his graduation, he joined Electronics and Telecommunications Research Institute (ETRI) where he had led Broadcasting Media Research Team and Realistic Broadcasting Research Team, and had worked in the MPEG-4/7 standardization related research areas. In 2001, he joined, as Assistant Professor in School of Engineering, the Information and Communications University (ICU) in Taejon, Korea. Since 2009, he is Associate Professor in Department of Electrical Engineering at KAIST, Daejeon, Korea. His research areas of interest include 2D/3D video coding, 3D video quality assessment, pattern recognition and machine learning, and video analysis and understanding.