Abstract
Streaming services spread rapidly. Among these services there are the linear TV, video library or program review system, while the online platform offering these contents is called mobile TV. A recommendation system may not only keep existing clients, but may also generate further turnover, should it introduce new content to the users. In this paper a recommendation system based on the Élő point calculation method is addressed. It is detailed how the programs should be grouped into different dimensions and what type of categories should be considered. Further, the idea of punch cards is introduced. Besides, the user profiles are set. The match system introduced by Élő is applied to the present situation. The system is introduced at a local mobile TV provider with 20,000 users.
1 Introduction
Due to the rapid expansion of the media streaming services, recommendation systems gain more attention. With the help of recommendation systems, the target is to provide the users with an immediate choice, in other words to minimize the time for selecting the next content and instead to maximize the entertainment time. Thus, the client is kept satisfied and additional turnover may also be generated. For this reason, it is also of high importance to increase the reliability of the applications. Therefore, a kind of competition has developed between the developers in the development and improvement of the recommendation systems.
The interaction and delivery requirements for mobile TV have long been examined, for example Knoche and McCarthy [1]. They compared the features of traditional TV with mobile TV services and outlined the design requirements for an imagined mobile TV interface as well as investigated early bandwidth requirements. Gao et al. [2] proposed an Internet Video (IV) recommendation system working for Live TV Programs. They extended the Electronic Program Guide (EPG) information with results of various searches of the Internet, which is compared later with key frames of the live TV programs to enhance the understanding of the TV broadcasts. Besides, profiles of the subscribers are built on historical data to improve the system. Finally, they also disposed some sort of recommendation strategies.
In an article of Krishnappa et al. [3] the user behavior of requesting videos from the top of the related list is used to improve the performance of YouTube caches. They recommend that local caches reorder the related lists associated with YouTube videos, presenting the cached content above noncached content. According to their findings, the position on the short lists really is the selection criterion more dominant than the content itself. According to the opinion of Li et al. [4] video streaming services heavily depend on the video recommender system. While most existing recommender systems compute video relevance based on users implicit feedback, they propose a deep convolutional neural network to alleviate the problem of a newly added video, i.e., to bootstrap the video relevance score with very few user behavior. Later, Agarwal et al. [5] stated that the main problem with recommending systems on linear TV is the absence of explicit ratings from the user. They assign different weightage to both frequency and duration of each user-show interaction pairs in order to provide suitable recommendations. Recently, Cardoso and Abreu [6] accept that nowadays, recommendation systems are everywhere, ranging from e-commerce sites to the media streaming services. But unfortunately, in most pay-TV platforms, these recommendation systems do not usually achieve a suitable level of effectiveness. They suggested some reasons for that issue and also proposed a personalization system, and described a prototype to overcome some shortcomings.
The Élő rating system [7] is a method for calculating the relative skill levels of players in zero-sum games. It was introduced by Árpád Élő in 1978 for chess. Here, instead of points awarded on the basis of subjective ratios based on the ranking of competitions a statistical method is applied. For example, in golf, some competitions are prioritized over others; in other words the winner of a certain tournament can score up to five times as many points in the individual evaluation as the winner of another tournament. The statistical method instead uses a model in which the result of each match is determined by a variable representing the playing power of each player up to that point. According to the central assumption, if the player wins a match, they are assumed to have performed at a higher level than their opponent. Moreover, beating a stronger player is a good result that has a more positive effect on the winner's playing power variable than if he had won against a relatively weaker player. Conversely, if the player loses, they are assumed to have performed at a lower level. Moreover, if someone is outplayed by an opponent with a lower playing strength, it has a negative effect on their own playing strength variable. Further, in case of a draw, the two players are assumed to have performed at nearly the same level.
Zak [8] applied Élő method to analyze the performance advantage of traveling. The study corresponds to international tournament chess. The idea behind was that many individuals travel between countries as part of their professional routines. According to the findings there is enhanced performance among players who were competing outside of their home countries. This finding was robust to additional controls for example age, sex, and skill momentum or game practice, and to the inclusion of individual or country fixed effects. This advantage, i.e., the increase in game outcome, suggests that traveling has a positive effect on performance.
Élő rating method is applied for tennis match predictions, see for example Angelini et al. [9]. Their method estimates the strength of a player based on career historical data, where the standard Élő updating is additionally weighted according to the scoreline of the players' last match, and it is also considered how the last victory (defeat) was achieved by the player. Earlier, Balakrishnan and Chopra [10] modified the Élő system and created an adaptive pairwise preferences and latent factor models. There, the user is continuously asked to provide additional information to the system by entering feedback to a sequence of pairwise preference questions. Then, based on the new response information the user parameters are updated, and subsequent questions are chosen adaptively as further feedback. They operate a Bayesian framework and the choice of questions is based on an information gain criterion.
Élő method has been adopted in other contexts including team sports. Élő method is used for football ranking procedures; see for example the FIFA/Coca-Cola World Ranking [11] and the World Football Élő Rating [12]. These are in use as alternative approaches capturing an international team's expected playing strength. The most recent FIFA/Coca-Cola World Ranking is available since August 2018 [11]. This latest version relies on adding points won for a game to the previous points rather than averaging game points over a given time period as in the previous versions. The calculated added points are partially determined by the relative strength of the two opponents, including the logical expectation that teams higher in the ranking should fare better against teams lower in the ranking.
For American football, i.e., for the National Football League (NFL) teams, the Élő method was applied by Silver [13]. He developed a simulator program that plays out the NFL schedule thousands of times and projects a team's likelihood of making the playoffs, based on a team's record up to that point in time, its Elo rating, its remaining schedule and the NFL's various tiebreaker rules. Later for basketball, for the National Basketball Association (NBA) teams Silver and Fischer-Baum [14] also applied the Élő method to investigate the complete history of the NBA. The project was ongoing until the summer of 2023. This NBA Élő database served as a necessary input of Salaga et al. [15] who demonstrated that betting market outcomes are a statistically significant and economically relevant driver of local market television viewership. They draw some connections between sports wagering and viewership in less-popular games and when the local market team is expected to perform poorly.
Further considerations may be gained from solutions of theoretically complex problems, for example Frits and Bertók [16] and Ochieng et al. [17]. Fuzzy logic considerations by Jancskár et al. [18] and Pusztai et al. [19], or uncertainty issues by Pusztai et al. [20], may also motivate future analysis of the topic.
2 Proposed method
2.1 Schematic process of the recommendation system
The proposed schematic process of the recommendation system is depicted in Fig. 1. First, a punch card system is developed which corresponds to the classification and evaluation of the programs. Behavioral habits of the users are grouped into the following four dimensions: stimulus, show-time, time slot, subscription type. Within the dimensions there are various categories. For example, the following categories are available within the time slot dimension: between 0 and 6, between 6 and 10, between 8 and 12, between 10 and 14, between 12 and 16, between 14 and 20, between 18 and 24. There is an intentional overlap between these categories. The dimensions serve as basis of the evaluation system. The evaluation is done in accordance to the points calculated by the Élő rating method. Dimensions serve as the leagues within the Élő rating method, while the categories serve as the players. Leagues do not play against each other, but within a league every player has a match against every other player. The categories in each dimension are filled with scores from existing data and from viewing data, and are also continuously updated in real time. These values determine the relevance values of the categories within each dimension. Since the current system stores further information about the users, this is used for the cold start, which can also be considered as a kind of startup initialization.
Proposed schematic process of the recommendation system
Citation: Pollack Periodica 19, 3; 10.1556/606.2024.01063
2.2 Dimensions
The following four dimensions are used within the systems: stimulus, show-time, time slot, and subscription type. The stimulus can be gained from Meta-data of the programs, and they represent the group categories of the programs. The show-time dimension can be determined based on the viewing habits of the users, for example the user favours half an hour programs, or another user prefers films. Recommendations can be easily separated by the various time slots, since different habits correspond to different time slots usually. In other words, for the recommendation system different items will be considered to be relevant, thus different content will be recommended in different time slots. Further important information can be determined based on the user subscriptions. It is of outmost importance that content falling out of the subscription range should never be recommended to the given user. The various dimensions are considered with different weighting factors (Table 1). Evaluation of the dimensions is done separately.
Weighting of the dimensions
| id | Dimension | Weight |
| 1. | stimulus | 0.3 |
| 2. | show-time | 0.2 |
| 3. | time slot | 0.4 |
| 4. | subscription | 0.1 |
Within the dimensions categories are defined. For example, within the stimulus dimension the following categories are defined: entertainment, information, inspiration, social interaction, fantasy and adventure, relaxation. It is worth mentioning that should a film belong to the action category according to the Internet Movie DataBase (IMDB), then within the proposed system the film will belong to both the inspiration, and the fantasy and adventure categories. This is important to blur the sharp edges of each category and thus the next recommendation will not only contain content from only one category, but already films of two categories can be offered. The overlapping is intended for the categories of the other dimensions also. For example, should a user watch a program from 19 o'clock then both the afternoon and the evening contents will appear within the recommended section.
2.3 Punch cards
When the dimensions are set, the first task is to prepare the punch cards. Punch cards are defined when a program is recorded into the EPG. The lines of the punch cards represent the dimensions, while the columns of the punch cards represent the different categories of the various dimensions. During the preparation of the punch card, the IMDB categories included in the metadata received for each program are used to select the stimuli to be punched. Where a specific feature is representative, there the punch card is punched (✓), all other fields remain unpunched (x). The main role of the punch cards during the calculations is to determine which dimension elements are considered to be relevant, in other words winner, or non-relevant, aka looser. The punch card of the program is matched with the user's profile; where the user's points are visible, it is considered to be a won match, and elsewhere, i.e., where the user's points are not visible, it is considered to be a lost match - this is where the punch card's idea is originated. Table 2 illustrates an example for a film, where each dimension has its own punch card. The title of the program is Raiders of the Lost Ark [21].
Punch cards of the dimensions (A: stimulus, B: show-time, C: time slot, D: subscription) and overall punch card of the film Raiders of the Lost Ark (E)
| A. Stimulus dimension according to the IMDB category | ||||||
| 1–1 | 1–2 | 1–3 | 1–4 | 1–5 | 1–6 | 1–7 |
| Entertainment | Information | Inspiration | Social interaction | Fantasy and adventure | Relaxation | |
| ✓ | × | × | × | ✓ | × | |
| B. Show-time dimension: 1 h 55 min | ||||||
| 2–1 | 2–2 | 2–3 | 2–4 | 2–5 | 2–6 | 2–7 |
| 0–30 m | 20–60 m | 45–90 m | 60–120 m | 120 m + | ||
| × | × | × | ✓ | × | ||
| C. Time slot dimension: starts at 19:00 | ||||||
| 3–1 | 3–2 | 3–3 | 3–4 | 3–5 | 3–6 | 3–7 |
| 00:00–06:00 | 06:00–10:00 | 08:00–12:00 | 10:00–14:00 | 12:00–16:00 | 14:00–20:00 | 18:00–24:00 |
| × | × | × | × | × | ✓ | ✓ |
| D. Subscription dimension: available on HBO2 channel | ||||||
| 4–1 | 4–2 | 4–3 | 4–4 | 4–5 | 4–6 | 4–7 |
| Basic | Extra | Premium | Film | Kids | International | Music |
| × | × | ✓ | ✓ | × | × | × |
| E. Punch card of the film | ||||||
| ✓ | 1–2 | 1–3 | 1–4 | ✓ | 1–6 | |
| 2–1 | 2–2 | 2–3 | ✓ | 2–5 | ||
| 3–1 | 3–2 | 3–3 | 3–4 | 3–5 | ✓ | ✓ |
| 4–1 | 4–2 | ✓ | ✓ | 4–5 | 4–6 | 4–7 |
2.4 User profile
Tables containing the dimension are created for each user to match them with the punch cards of the programs. Table 3 gives the Élő points of the user that corresponds to the dimension-categories. The rows of the table correspond to the dimensions. The punch cards of the programs are matched with this user profile and the relevance of each category is examined based on the visibility of the points. Initially, each category has the value of 2,000 points, which is updated regularly. In the first dimension, there are 6 categories, and thus the sum of the category points is 12,000. In the second dimension, there are only 5 categories, and thus the sum of the category points is 10,000. The category points are recalculated, but these overall values do not change when the user profile is updated, i.e., the sum of the category points of the first dimension remains 12,000, while for the second dimension it remains 10,000.
Élő points of a user corresponding to the dimensions and their categories
| 2,500 | 2,200 | 1,200 | 1,900 | 2,200 | 2,000 | |
| 1,600 | 1,800 | 1,600 | 3,000 | 2,000 | ||
| 1,000 | 2,200 | 1,000 | 2,200 | 2,400 | 2,600 | 2,600 |
| 1,400 | 1,600 | 3,000 | 2,800 | 1,600 | 1,800 | 1,800 |
2.5 Recommendation block
When setting up the user specific recommendation block, the first step of the selection of the programs to be recommended is matching the user's profile with the program's punch card. This shows which categories are relevant within the dimensions. For the example, cells of Table 4 have bold and italic fonts, where the category of the given dimension is relevant, and cells have normal fonts, where the category is not relevant.
Matching the user's profile and the punch card of the film Raiders of the Lost Ark; the bold and italic font indicate cells that are relevant, other cells are not relevant
| 2,500 | 2,200 | 1,200 | 1,900 | 2,200 | 2,000 | |
| 1,600 | 1,800 | 1,600 | 3,000 | 2,000 | ||
| 1,000 | 2,200 | 1,000 | 2,200 | 2,400 | 2,600 | 2,600 |
| 1,400 | 1,600 | 3,000 | 2,800 | 1,600 | 1,800 | 1,800 |
The sigmoid logistic function in the equation well represents the probability change as a function of the point difference. The constant 400 is standard within the Élő method. The purpose of the formula is to create a probability model based on the results of matches between players. The two ends of the S-shaped curve have low and high values, while the middle part shows a more uniform change. This arrangement means that the change in probability between players slows down as the difference in point's increases or decreases.
2.6 Update
When a user has watched a program for more than a certain period of time and the program is either finished or the user switched to another channel, then the points within the user's profile are updated. Based on practical experience, this time limit is set to 5 min, and to 50% of the show-time of the program. Note that in case of live shows, when calculating the 50% of the show-time, it should also be considered whether the user watched the program in multiple fragments, in other words each viewing is measured and summed up for each program. This is to overcome the situation when advertisement blocks are diverting the users from the given program. The id within the program guide supports this measurement. Nevertheless, in case of Video On Demand (VODs) it is not considered, i.e., each start of the VOD program is handled separately.
First, in accordance with the calculations of the Élő point system, the matches between each player within the leagues are determined. Obviously, no one will play a match against itself. Should the category be relevant then the cell of the table will receive the value ‘1,’ should the category be not relevant then this value is set to ‘0’. Should this intersection be the same, i.e., both categories be relevant, and neither of them be relevant, then these cells receive the ‘0.5’ value. In other words, the category matrix of every dimension is determined. Within the dimension the possibilities of each category pair is calculated. The given pairing can be calculated by selecting a category for a given row and column of the matrix. The row and column category value indicates the current score of the categories in the selected pairing. Referring back to the previous example, the film Raiders of the Lost Ark, Table 5 illustrates this calculation. There the match between the Entertainment and Information categories is ‘1’ since the column is the Entertainment, which is relevant to the program, while Information is not relevant. Similarly, the match between Information and Entertainment categories received ‘0’ value. Note that this table can be generated mirrored to the main diagonal.
Matches between the players, i.e., categories of the stimulus dimension of the film Raiders of the Lost Ark
| Entertainment | Information | Inspiration | Social interaction | Fantasy and adventure | Relaxation | |
| Entertainment | – | 1.0 | 1.0 | 1.0 | 0.5 | 1.0 |
| Information | 0 | – | 0.5 | 0.5 | 0 | 0.5 |
| Inspiration | 0 | 0.5 | – | 0.5 | 0 | 0.5 |
| Social interaction | 0 | 0.5 | 0.5 | – | 0 | 0.5 |
| Fantasy and adventure | 0 | 1.0 | 1.0 | 1.0 | – | 1.0 |
| Relaxation | 0 | 0 | 0.5 | 0.5 | 0 | – |
Based on the matches between the categories, the points of the user should be updated according to Eqs (6)-(8). Further, these values are averaged (Eq. (9)). The constant value 16 is applied in Eq. (8); it is adjusted to the standardized value 400 in the probability formula. Other values, for example 24, could also be used, depending on the dynamic of the system. Note that during this calculation the sum of the points of the categories within a dimension remain unchanged. Further note that these equations are similar to the percentage values of Eqs (1)-(2), nevertheless here the equations refer to probability values.
Updated user's points after the film Raiders of the Lost Ark
| Entertainment | Information | Inspiration | Social interaction | Fantasy and adventure | Relaxation | Updated points | |
| Entertainment | – | 2,502 | 2,500 | 2,500 | 2,494 | 2,501 | 2,500 |
| Information | 2,198 | – | 2,192 | 2,194 | 2,192 | 2,196 | 2,194 |
| Inspiration | 1,200 | 1,208 | – | 1,208 | 1,200 | 1,208 | 1,205 |
| Social interaction | 1,900 | 1,906 | 1,892 | – | 1,898 | 1,902 | 1,899 |
| Fantasy and adventure | 2,206 | 2,208 | 2,200 | 2,202 | – | 2,204 | 2,204 |
| Relaxation | 1,999 | 2,004 | 1,992 | 1,998 | 1,996 | – | 1,998 |
2.7 Start
When the mobile TV user interface is started, it is important to identify the user. To determine whether the user is a new or an already existing one, historical records are considered. Should there be no available historical records; the initial values are set to 2,000 points as mentioned earlier. Should some initial profile be built based on the contract or from somewhere else, those modified values are included in the profile. Based on the available points, an initial recommendation section already appears.
3 Discussion
The current recommendation system is introduced at a TV provider with 20,000 users. The recommendation content consists of live shows, video library as well as VODs. The subscription base corresponds to an approximately 1,000 users during the evening hours. The recommendation section consists of 10 programs. It is measured, which program is started and for how long is the program watched in percentage. For the below measurement 1,000 recommendation sections were selected as samples. Figure 2 shows that should the first content be of no interest for the user, then the further programs are neglected and the recommendation section is left, the next program is more likely to be selected by the users themselves. Obviously, user interface ergonomics' issues are also of high importance which are not scope of the current paper.
Started program: 1 through 10 within the recommendation section and from outside
Citation: Pollack Periodica 19, 3; 10.1556/606.2024.01063
Table 7 summarizes the behavior patterns of the users in terms of show-time or duration of the program. It is important to mention that in case of live program show-time contains one or more advertisement blocks, which is usually skipped by the users. This may also result in a situation where 70% of the overall show-time equals to the pure duration of the program.
Behavior pattern of the users in terms of show-time
| Show-time | Number of users |
| 100%–80% | 474 |
| 80%–60% | 378 |
| 60%–40% | 45 |
| 40%–20% | 32 |
| 20%–0% | 71 |
References
- [1]↑
H. Knoche and J. D. McCarthy, “Design requirements for mobile TV,” in Proceedings of the 7th International Conference on Human Computer Interaction with Mobile Devices & Services, New York, NY, USA, Sep. 19, 2005, pp. 69–76.
- [2]↑
S. Gao, D. Zhang, H. Zhang, J. Liao, C. Huang, Y. Zhang, and J. Guo, “VecLP: A realtime video recommendation system for live TV programs,” in Proceedings of the AAAI Conference on Artificial Intelligence, Austin, Texas USA, January 25–30, 2015, vol. 29, no. 1, pp. 4274–4275.
- [3]↑
D. K. Krishnappa, M. Zink, C. Griwodz, and P. Halvorsen, “Cache-centric video recommendation,” ACM Trans. Multimedia Comput. Commun. Appl., vol. 11, no. 4, pp. 1–20, 2015.
- [4]↑
Y. Li, H. Wang, H. Liu, and B. Chen, “A study on content-based video recommendation,” in 2017 IEEE International Conference on Image Processing, Beijing, China, Sep. 17–20, 2017, pp. 4581–4585.
- [5]↑
A. Agarwal, S. Das, J. Das, and S. Majumder, “A framework for linear TV recommendation by leveraging implicit feedback,” in Proceedings of Computational Science and Technology, Kota Kinabalu, Malaysia, August 29–30, 2018, Lecture Notes in Electrical Engineering, vol. 481, 2019, pp. 155–164.
- [6]↑
B. Cardoso and J. F. De Abreu, “TV personalization: blending linear and on-demand content in the living room,” Int. J. Entertainment Technol. Manage., vol. 1, no. 2, pp. 162–177, 2021.
- [8]↑
U. Zak, “The performance advantage of traveling,” J. Econ. Psychol., vol. 87, 2021, Art no. 102431.
- [9]↑
G. Angelini, V. Candila, and L. De Angelis, “Weighted Elo rating for tennis match predictions,” Eur. J. Oper. Res., vol. 297, no. 1, pp. 120–132, 2022.
- [10]↑
S. Balakrishnan and S. Chopra, “Two of a kind or the ratings game? Adaptive pairwise preferences and latent factor models,” Front Comput. Sci., vol. 6, no. 2, pp. 197–208, 2012.
- [11]↑
Revision of the FIFA/Coca-Cola World Ranking, FIFA. [Online]. Available: https://www.fifa.com/fifa-world-ranking/procedure-men. Accessed: Jan. 22, 2024.
- [12]↑
World football Elo ratings [Online]. Available: https://www.eloratings.net/. Accessed: Jan. 22, 2024.
- [13]↑
N. Silver, “Introducing NFL Elo Ratings.” [Online]. Available: https://fivethirtyeight.com/features/introducing-nfl-elo-ratings/. Accessed: Jan. 22, 2024.
- [14]↑
N. Silver and R. Fischer-Baum, “How we calculate NBA Elo ratings.” [Online]. Available: https://fivethirtyeight.com/features/how-we-calculate-nba-elo-ratings/. Accessed: Jan. 22, 2024.
- [15]↑
S. Salaga, S. Tainsky, and M. Mondello, “Betting market outcomes and NBA television viewership,” J. Sport Manage., vol. 34, no. 2, pp. 161–172, 2020.
- [16]↑
M. Frits and B. Bertok, “Routing and scheduling field service operation by P-graph,” Comput. Oper. Res., vol. 136, 2021, Art no. 105472.
- [17]↑
P. J. Ochieng, A. London, and M. Krész, “A forward-looking approach to compare Ranking methods for sports,” Information, vol. 13, no. 5, 2022, Art no. 232.
- [18]↑
I. Jancskar, Z. Sári, A. Schiffer, and G. Várady, “Phase plane tuning of fuzzy controller for 1 DoF helicopter model,” Pollack Period., vol. 10, no. 2, pp. 3–15, 2015.
- [19]↑
L. Pusztai, B. Kocsi, and I. Budai, “Making engineering projects more thoughtful with the use of fuzzy value-based project planning,” Pollack Period., vol. 14, no. 1, pp. 25–34, 2019.
- [20]↑
L. Pusztai, B. Kocsi, I. Budai, and L. Nagy, “Investigation of a production process under uncertainty,” Pollack Period., vol. 15, no. 2, pp. 49–59, 2020.
- [21]↑
IMDB record of the movie Raiders of the Lost Ark. [Online]. Available: https://www.imdb.com/title/tt0082971/. Accessed: Jan. 22, 2024.
