Keywords

1 Introduction

Watching online videos is a big entertainment in people’s lives these days. The huge success of YouTube and some other online video websites shows people have great interest in watching these fun, UGC videos. A number of online video websites are founded during the last decade, and to attract viewers and seek success, new features are introduced.

In 2007, Justin Kan founded Justin.tv. He first broadcasted his life 24/7 in this website, which led to the success of Justin.tv and even popularized the term lifecasting [1]. Since then, live-streaming websites arose all around the world and gained their popularity day by day.

Twitch.tv, which is an online video website focuses on live-streaming game videos also founded by Justin Kan, now ranks 77th popular among all websites worldwide by Alexa.

In China, hundreds to thousands new live-streaming websites and applications emerged in the last two to three years, making year 2016 “the Year of live-streaming” in China. Any events from the streamer’s life to outdoor activities to commercial events can be broadcasted. Right after the 2016 summer Olympics in Brazil, the famous Chinese swimming athlete Yuanhui Fu’s live-streaming attracted over ten million viewers online.

Seeing the heat of live-streaming websites, the academic world is interested in this phenomenon and many studies are done to seek the understanding of the user behavior, the feeling and benefits of watching these videos.

Kaytoue et al. [2] did a study on video game live streaming on Twitch.tv. He researched the audience, stream and streamer characteristics and studied how to predict the popularity of streams. Deng et al. [3] made a detailed research on the games played on Twitch, their features, channel and viewer distribution among different games and the impact of tournaments. Pires and Simon [4] compared the stream features of Twitch and YouTube. Hamilton [5] investigated how live streaming fosters participation and community. Nascimento et al. [6] and Edge [7] studied the community of live-streaming websites.

Though live-streaming websites are new, virtual gifts already exist for some time. Goode et al. [8] studied the virtual gifts in virtual world and concluded sending gifts enhances users’ social status. Yang et al. [9] researched the connection between virtual gifts and interpersonal influence. Greenberg [10] put forward a possible technique for introducing virtual gift sending interaction into online live car racing games.

Despite the fact there are lots of research on the above topics, as far as we know, little has been done on the gift-sending interaction on live-streaming websites. Tough in fact virtual gifts have been the top earner in Chinese live-streaming markets [11].

In this paper, we present our findings on the website and the gift-sending interaction, an activity viewers interact with streamers and a way these websites earn profits. We did our experiment using the data of one of the most popular live-streaming websites in China-Douyu.com.

The rest of the paper is organized as follow. In Sect. 2, we introduce the background of Douyu.com, the origin of gift-sending interaction and the dataset we use. We get some basic information of Douyu.com in Sect. 3 and investigate gift-sending interaction in Sect. 4. In Sect. 5, we discuss future research aspects. The conclusion is in Sect. 6.

2 Background and Research Data

Founded in 2014, Douyu.com is now one of the biggest online live-streaming websites in China, ranks 288th popular in the world and 42nd popular in China by Alexa rank. It has on average over 7 million UV and over 66 million PV a day. It is a comprehensive live-streaming website, the broadcasted content includes computer games, outdoor activities, people’s daily lives, singing and dancing, popular stars and commercial events. Among them computer games occupy a very big proportion.

People who broadcast are called streamers or broadcasters, every streamer has a single web page in this website named channel, those who watch these channels are viewers.

A typical page of a channel on Douyu.com is shown in Fig. 1. This page has the streamer’s information, the content he is broadcasting, and gifts viewers can buy. The right side text part is viewers’ chat messages. Most live-streaming websites’ channel pages are alike. For those websites do not have gift functions like Twitch, the only difference is Twitch pages do not have the bottom gifts options and thus do not have gifts information on the screen.

Fig. 1.
figure 1

A typical channel page in Douyu.com

Full size image

Now it naturally comes to the question: How does gift-sending interaction originate?

Virtual gifts are used by some live-streaming websites to gain profit. When a gift is bought and sent, the streamer and the website share its profit. Gifts all have their prices, usually ranging from 0.1 RMB to 500 RMB (we leave out gifts which users can get for free and the money unit in this paper is RMB).

Virtual gifts are extension to chat messages in live-streaming websites. When a gift is sent, the system lists this gift’s information on the right text part of the channel. All viewers in this channel can see this. If the gift has a high value, for example, 100 RMB or higher, the system even projects the gift’s icon directly on the left screen part. The higher the gift value, the more complex and vivid the icon.

This gives viewers a new way they interact with streamers. In traditional live-streaming websites like Twitch, there are few ways to distinguish viewers. But with these gifts, rich viewers accumulate their “fame” and level in this channel. The system lists top gift senders’ information in the text part and welcomes them when they enter the channel, which makes them outstands, both in the streamer’s eyes and other viewers’ eyes.

Dataset.

We crawled data of Douyu.com from 2016 Dec.17 to Dec.30. The dataset contains two kinds of data. One is about the global information of this website. We crawled the whole website’s opening channels’ data, including channel number, viewer number and channel type, every 10 min for a 14-day period. The other is one hundred popular channels’ detail data for 14 days. In every channel, every gift type id, the sender’s id, the timestamp (detail to minute) are recorded.

3 Basic Information About Douyu.com

We decide to get some basic information about Douyu.com before digging into the gift-sending interaction on it. In this section, we use the dataset of a fourteen-day continuous data of all the channels’ information on Douyu.com to do the study.

The Channel Number.

Here we conclude some basic information of the channel number in this website in Table 1.

Table 1. Channel number information
Full size table

Though on average there are 35989 unique opening channels in a day on Douyu.com, the opening channel number in any snapshot is not that high. It varies between 1939 and 11538. On average about 5984 channels are open in the same time.

Opening Pattern.

In Fig. 2, we show the relation about channels’ opening pattern. We see about 63% of 149424 channels open no more than one day a week on average, only 14% channels open at least half the days in dataset. Most channels do not broadcast streams frequently. But there are about 4% of them open every day.

Fig. 2.
figure 2

CDF of channels by opening days

Full size image

Duration of Streams.

In this part, we focus on the streams in the channels, instead of channels themselves. If a channel opens a piece of continuous time then stops, for example an hour, we call it a stream or a session. A single channel can have several streams in a single day as long as the streamer takes break during these streams. Using streams here is better than channel itself because we can observe how streamers choose to open and stop their live-streaming videos.

The distribution of stream duration is in Fig. 3. There are 686086 streams in the two-week dataset. Since our snapshot is every 10 min, whenever we see a channel in the dataset, we take it as having been opened for 10 min.

Fig. 3.
figure 3

CDF of streams by duration

Full size image

From Fig. 3 we can see the median of stream duration is 90 min. It is longer than the 45 min median time of stream in Twitch [2]. That’s because unlike Twitch which focuses on games, Douyu.com is a more comprehensive website. Many channels broadcast already-made videos or big events, that raises the duration of streams.

Besides the 45 min median, over 70% streams are shorter than 200 min. The lasting time of most streams is no longer than 4 h, because the streamers need rest.

Changing Pattern of Channel and Viewer Number During a Day.

From Table 1 we learn the channel number in Douyu.com varies during the time. So here we plot channel number and viewer numberFootnote 1 to see this pattern. Fourteen-day data is averaged into 24 h in Fig. 4.

Fig. 4.
figure 4

Channel number changing pattern

Full size image

The opening channel number shows a clear varying pattern in Fig. 4. It becomes lowest in early morning around 7 am, then gets higher. There is a small peak in around 16 pm in the afternoon, a small drop follows this peak during dinner time then the number becomes higher again, reaching the peak in 9–10 pm. The total viewer number information is in Fig. 5. It is in consistence with this varying pattern, the only difference is the small peak in the afternoon is a little earlier than streamers do.

Fig. 5.
figure 5

Viewer number changing pattern

Full size image

Kaytoue et al. [2] also found the viewer count in a day on Twitch has two peaks. He owed these two peaks to different users from Europe and America. We agree with his findings on Twitch since its two peaks are more alike. But considering Douyu.com’s viewers are mainly from mainland China and the two peaks are not quite the same in quantity, we think the two peaks in Douyu.com maybe just because many people have free time in the afternoon in China.

Channel Types, Viewer Types and Their Proportions.

Douyu.com labels different channels according to their contents. There are five big categories defined by Douyu.com: PC Games, Mobile Games, Entertainment & Activities (including sing, dancing, music, outdoor activities, comics and so on), Technology and Art & Sports (including educational videos, old movies, sports and so on).

These five categories occupy about 80% of the total videos in Douyu.com. Their proportion is in Table 2. Under these categories, there are smaller types of channels, like League of legends or DOTA2 in PC Games, music or fashion in Entertainment & Activities. We also list top 10 popular types and their proportions in Table 3.

Table 2. Channel categories and proportion
Full size table
Table 3. Top 10 channel Types (*means game)
Full size table

In Table 2, PC Game channels do occupy a big proportion in this website, over 50%. This is one big feature how Douyu.com earns its original fame. Now many other contents like Entertainment & Activities and Art & Sports have their cakes too.

In the top 10 channel types by channel number, seven of them are games, among them Glory of the king is a mobile game. Movie & TV and Sports are from Art & Sports. Beauties is from Entertainment & Activities. This definitely shows Douyu.com is a comprehensive live-streaming website.

In Table 3 we also show the top 10 channel types that attract the most viewers. Six of them are in the top 10 channel types by channel number. Despite the fact the channel numbers of Outdoor Activities, DOTA2 and Hearthstone rank 11th, 12th, 13th, their viewer numbers are in the top 10. Mobile game Onmyoji is a new game released on 2016 September. It doesn’t have many channels, only ranked 26th, but have a big number of viewers because of its popularity.

4 Gift-Sending Interaction Study

Unlike Twitch, which uses subscription, stickers and advertisements to help streamers and website earn money, gifts are widely used in many Chinese live-streaming websites to gain profit. In 2016, Greenberg [10] proposed a technique that viewers can send gifts as a game resource in online car racing games. This proposed technique has some similarity, but not quite the same as the widely used gift service in Chinese live-streaming websites. Up to now, little experiment about virtual gifts is done in real live-streaming websites.

In this part, we study the gift-sending interaction in Douyu.com. We propose several research questions, demonstrate their research methods and results in this section.

Q1: How do viewers choose between different gifts?

In Sect. 2 we have described the origin and price features of gifts. Virtual gifts are used by live-streaming websites to gain profit. When a virtual gift is bought and sent to a channel, its value is shared by the website and the streamer.

Here we want to answer the question about how do people choose between cheap or expensive gifts? What is the consumed amount of each gift type? What is the total value of these gifts?

To answer that, we listed some basic information of gifts in Table 4. Douyu.com has a number of gift typesFootnote 2, some are available for all channels and some are special designed for typical channel types. Here for simplicity we only listed five specific gifts that are common for all channels, since they occupied 97.2% of the total gift value in the dataset.

Table 4. Gifts and their purchase details
Full size table

From Table 4 we can observe the distributions of gifts and senders are very skewed, even severer than the well-known “80/20” rule. The consumed number of high-value gifts is much smaller than that for cheap-value gifts, but they occupy a high proportion of the total gift value. And it is noteworthy about 80.2% of the total gift value are generated by only 2.7% of the total consumers. When sending gifts, the majority viewers only buy cheap gifts while the majority profit comes from minority viewers.

Q2: How does the gift-sending interaction look like on a high level for channels?

After studying gifts on a viewer level, now we want to find out the high-level result of gift-sending interaction.

In the dataset we have 100 channels randomly chosen from top 1000 channels. We count their gift value sum to see this high level pattern. Result is in Fig. 6.

Fig. 6.
figure 6

100 channels’ gift value sum

Full size image

The Y-axis in Fig. 6 is the gift value sum of each channel and the X-axis is channels’ rank by money. Only 86 channels get gifts out of the chosen 100. We also wrote the 1st, 11th, 21st, 31st…81st channel’s gift value sum in Fig. 6. A power-law exists, indicating several most popular channels have the most valuable gifts, others share little.

Q3: What’s the relationship between viewer number and gift value sum of a channel?

It is easy to think that both viewer number and the time they spend on a channel may have impacts on the final value of gifts they buy. So here we use “viewer appearance number” instead of average viewer number. All the viewer number of a channel in every 10-min snapshot are summed as the final viewer appearance number of a channel.

We plot gift value sum and viewer appearance number of every channel in Fig. 7. The 14 channels whose total gift value is 0 are removed and 86 channel’s data are left.

Fig. 7.
figure 7

Channel gift value sum and viewer appearance number

Full size image

It is easy to observe that though the absolute relation is volatile, in general gift value sum has increasing nature with the increment of the viewer appearance number.

We do a linear regression of log10(gift value sum) on log10(viewer appearance number) and the result is in Table 5. The coefficient is 0.6421, p < 0.001, generally speaking, the logarithmic value of gift value sum is positively correlated with the logarithmic value of viewer appearance number, so is their original value.

Table 5. Regression information of log10(gift value sum) on log10(viewer appearance number)
Full size table

Generally speaking, if a channel attracts more viewers and let them spend more time on itself, it gets more gifts and earns more. Popularity is important for a streamer like any other industry. More fans, more benefits.

Q4: Will other viewers’ gift-sending behavior in the same channel stimulate viewers to send gifts?

We compute probability and conditional probability to study viewers’ behavior in the 100 channels. The time unit is one minute. The probability for a viewer \( v_{i} \) sending gifts in a channel c is:

$$ {\text{P(}}v_{i,c} )= \frac{{n_{i,send}^{c} }}{{n_{open}^{c} }} $$
(1)

where \( n_{open}^{c} \) is the total number of minutes channel c opens and \( n_{i, send}^{c} \) is the total number of minutes viewer \( v_{i} \) sends at least one gift in channel c. Conditional probability for a viewer \( v_{i} \) sending gifts in a channel c is:

$$ {\text{P}}\left( {v_{i,c} |v_{j \ne i,c} } \right) = \frac{{n_{{i,send^{{\prime }} }}^{c} }}{{n_{j,send}^{c} }} $$
(2)

where \( n_{j,send}^{c} \) is the total number of minutes having other viewers \( v_{j,j \ne i} \) sending gifts, and \( n_{{i, send^{{\prime }} }}^{c} \) is the total number of minutes viewer \( v_{i} \) sends at least one gift in this minute later than other viewers do.

We then plot each viewer’s data in Fig. 8 where x-axis is \( {\text{P}}\left( {v_{i,c} } \right) \) and y-axis is \( {\text{P}}\left( {v_{i,c} |v_{j \ne i,c} } \right) \). We draw a point only when this viewer has sent a gift in this channel. 71.8% (20449 out of 28499) points are in the upper left part of Fig. 8, above y = x line, which indicates people’s motivation for sending gifts is positively correlated with others’ behavior. When seeing others sending gifts, they are more likely to do the same thing.

Fig. 8.
figure 8

Probability for sending gifts

Full size image

5 Discussion and Future Work

In this paper, we did research on live-streaming websites and the gift-sending interaction on it. Despite the fact we have a basic understanding of viewers and streamers’ behavior in this website now, there are some parts that worth discussion.

First, our dataset only contains two weeks’ data. This can give us a basic understanding about how viewers send gifts. But for more detailed research, a longer dataset period is better.

Second, live-streaming websites support Danmaku function, a function that projects users’ comment directly and synchronously on the play screen. Danmaku function is studied in traditional video websites [12, 13], but not in live-streaming websites as far as we know. This can be a future study focus too.

Third, in the gift-sending interaction, we do experiments about what facts will stimulate viewers to send gifts, we think there may be other facts that worth future work. For example, in every channel, a viewer can accumulate “fame” and their user levels when buying and sending gifts, the streamers usually orally thank these viewers to show their gratification. Will this oral interaction satisfy users and stimulate them into sending gifts? Will the accumulated “fame” encourage them sending more gifts in the future? Will viewers behave differently in different types of channels? We think these all need future work.

Another aspect we think need more future research is the topology of the user following pattern. Most live-streaming websites have following systems, through which a user can follow other viewers to get their information. We are interested in this following system and crawled about thirty million users’ following relationship data from Twich.tv in 2016 July. To our disappointment, the data shows the present relationship doesn’t show much bi-directional interaction in it. The bi-link proportion is very low, roughly 1.09%, comparted to twitter’s 22.1% [14]. Users in live-streaming websites like Twitch do not become friends. They just follow famous people. We think this topology phenomenon worth more research in the future.

6 Conclusion

In this paper, we present a study of live-streaming website and the gift-sending interaction in one famous website Douyu.com in China. To the best of our known, it is the first research that focuses on gift-sending interaction between streamers and viewers in real live-streaming website.

Our findings show the distributions of gifts and senders are very skewed in gift-sending phenomenon. Over 80% value of gifts are consumed by only 2.7% of viewers. And Several most popular channels get most gifts, others share little. Meanwhile, many facts have relations with the gift value sum. In general, if a channel attracts more viewers and let them spend more time on itself, it gets more gifts and earns more. Also, 71.8% viewers are likely to send gifts when they see others do, which shows stimulation is useful here.