The Emergence of Threads: The Birth of a New

Social Network
Peixian Zhang, Yupeng He, Ehsan-Ul Haq, Jiahui He, and Gareth Tyson
Hong Kong University of Science and Technology (Guangzhou)
Email: {pzhang041, yhe382}@[Link]
euhaq@[Link] jhe976@[Link] gtyson@[Link]

Abstract—Threads, a new microblogging platform from Meta, from Twitter to Mastodon, it did not change any topical
was launched in July 2023. In contrast to prior new platforms, preference of users [4]. We argue that the identification of
Threads was borne out of an existing parent platform, Instagram, such communication patterns is important for the subsequent
arXiv:2406.19277v1 [[Link]] 27 Jun 2024

for which all users must already possess an account. This offers
a unique opportunity to study platform evolution, to understand research on user modelling and discourse analysis [5].
how one existing platform can support the “birth” of another. Given the lack of research on Threads, here we present
With this in mind, this paper provides an initial exploration of the first analysis of early adopters of Threads, and compare
Threads, contrasting it with its parent, Instagram. We compare their discourse with those users on Instagram at the same
user behaviour within and across the two social media platforms, time. This analysis will provide a first benchmark of Threads
focusing on posting frequency, content preferences, and engage-
ment patterns. Utilising a temporal analysis framework, we iden- communication norms for future research. Specifically, we
tify consistent daily posting trends on the parent platform and answer the following research questions:
uncover contrasting behaviours when comparing intra-platform RQ1: What are the differences in posting behavior on Insta-
and cross-platform activities. Our findings reveal that Threads
engages more with political and AI-related topics, compared to
gram vs. Threads? (Section IV).
Instagram which focuses more on lifestyle and fashion topics. RQ2: What are the content differences between Instagram
Our analysis also shows that user activities align more closely on and Threads, at an overall platform level? (Section V).
weekends across both platforms. Engagement analysis suggests
that users prefer to post about topics that garner more likes and RQ3: For an individual user, what factors affect the change
that topic consistency is maintained when users transition from of topic in posts? Do these factors also impact users’ selection
Instagram to Threads. Our research provides insights into user of topics when they migrate from Instagram to Threads?
behaviour and offers a basis for future studies on Threads. (Section VI).
Index Terms—Instagram, Threads, User behaviour, Topic anal-
ysis Our analysis shows that Threads and Instagram have key
differences in the number of active users on weekdays during
the early days of Threads’ launch. In addition, there we find
I. I NTRODUCTION
a difference in topical focus, with topics like threads,
Meta’s new microblogging service, Threads, was launched twitter, and trump being more prevalent on Threads
on July 5th, 2023. It became the fastest-growing 100 million as compared to Instagram. Finally, we show that users who
consumer product within five days of its launch. In contrast consecutively write on similar topics have lower feedback on
to prior new social networks, Threads requires an Instagram their posts than those who write on different topics in their
account, thus enabling rapid social bootstrapping for Threads. consecutive posts.
Threads’ design allows users to follow selected or all of their
II. R ELATED W ORK
Instagram follow-network already on Threads. This relation-
ship between Threads and Instagram produces a distinct form We first discuss the prior work on multi-social platform
of attention migration attributed to the unique “parent-child”. comparisons and the platform migration of users.
This mass sign-up on Threads offers a unique perspective Multi-social network analysis. In order to conduct a com-
to characterise the discourse of early adopters on Threads. parative analysis of the same users’ activities across many
More importantly, it is possible to track these early adopters platforms, prior studies employ algorithms to establish con-
on the parent platform and offer a comparative analysis of nections between users’ accounts on other platforms [6].
users’ behaviors to the new platform. This is motivated by Additionally, they utilise biography websites [2] and examine
prior work that has revealed that users tend to follow varying self-mentions within the material across several platforms [7].
communication norms on different platforms. For example, These methods might bring different selection biases among
users share more negative sentiment and work-related content users who utilize biography websites or self-report their per-
on Twitter than on Instagram [1]. Similarly, Instagram users sonal information. In contrast, the direct connection between
tend to utilise profile images with lower smile scores compared Threads and Instagram obviates the need for intermediary
to Twitter users [2], and they share more content on weekends services, simplifying the process of cross-platform analysis.
than on weekdays [3]. In contrast, in the case of user migration We posit this offers a powerful ground truth dataset.
 There have been prior studies that focus on analyzing cross- extensive search range for early Threads adopters. For each
platform behaviours. Studies have found that users exhibit Threads account, we collect the 25 most recent threads posts
differences in self-description [3] and image usage [2] in with their details from that account (25 is the limit imposed by
profiles, alongside topic preferences [1] and activity time [3]. the API). From September 4th to September 13th, we further
Instagram, for instance, tends to be used more during leisure extend this data by snowballing the identification of additional
than work hours [3]. The engagement for similar content also users, and collecting their posts, based on the reposted threads
varies across different platforms [1, 8]. User activities also dif- (i.e. , users sharing other users’ content, similar to retweets on
fer in multi-social media platforms in terms of usage time [9]. Twitter). In total, there are 1,253,438 Threads accounts with
Moreover, there is a noticeable inclination among users to 4,716,626 posts. There are 582,459 (46.47%) users without
prioritize certain platforms over others when disseminating any threads, 543,522 (43.36%) users with fewer than 25
similar content [3, 8]. Based on these previous findings, we threads, and 127,457 (10.16%) users with 25 threads. We also
aim to determine whether Threads exhibits comparable activity find 212 threads created before the Threads launched (which
levels and topical preferences to its parent, Instagram. To the we exclude from the dataset). A possible explanation could
best of our knowledge, this is the first work to measure user be the presence of test accounts on Threads that Threads
activity on Threads. developers used to test the application’s functions before its
Platform Migration. Several prior works have studied the launch.
migration of users to different social media platforms. There 2) Instagram: Threads and Instagram share the same user-
are two types of migration: permanent migration and attention name, thus we search for all the usernames from the Threads
migration [10, 11]. [12] explore the push and pull factors that dataset on Instagram. We use the CrowdTangle API2 to retrieve
trigger migration on Reddit. There has also been prior work all corresponding Instagram accounts and their posts from May
studying attention migrations from Twitter to Mastodon [11], 5 to September 13, 2023. In total, we collect 683,168 accounts
and comparing the activities of the same users after migra- with 10,862,421 posts on Instagram based on the user list on
tion [4]. A key contribution of our work is the identification Threads. Note, we fail to collect any Instagram accounts that
of a third type of migration, which we term parent-to-child have their profiles set as private, or the accounts that have
migration. This is where an existing platform spawns a new changed their usernames.
service, as is the case with Threads, which was borne out For better comparative analysis with Threads, we split the
of Instagram (e.g. only Instagram account owners can access Instagram data into two parts: (i) Before Threads launch from
Threads, and Threads posts appear on Instagram timelines as May 4th to July 4th with 5,215,434 posts from 575,204 users;
a promotional tool). and (ii) After Threads launch from July 5th to Sept 9th with
To the best of our knowledge, the closest study to our own 5,646,987 posts from 587,941 users. This segmentation helps
is [13]. This work explored the characteristics differentiating compare users on two platforms after they start posting on
Twitter users who transition to Threads from those who stay Threads. There are 479,877 common users in the two datasets.
behind. However, this study does not investigate the specific Data pre-processing. We undertake several pre-processing
activities of the users on Threads, and its scope is limited to steps. On Instagram, the number of posts by each user is
a few thousand individuals. In our study, we utilize a dataset not normally distributed. The distribution of posts per user
comprising millions of users on Threads. Furthermore, all of is (µ = 9.6, median = 4, min = 1, max = 4, 942),
these previous investigations [4, 11, 13] focus on functionally (µ = 9, median = 4, min = 1, max = 4, 016) for before and
similar platforms. In our research, Threads and Instagram after launch data segment, respectively. Thus, we exclude any
represent distinct platforms that face a special migration with users with posts more than 0.99 percentile of number of posts
a novel online platform [10]. To the best of our knowledge, per user distribution. This covers 83 accounts for Instagram
this represents the first study to amalgamate data from both before Threads launched and 89 after Threads launched. We
Instagram and Threads for a comparative analysis of user also exclude the same users on Threads to ensure that the
engagement. comparison is conducted on the same set of users. For RQ 1,
this gives us 7,856,931 posts from 819,043 users on Threads,
III. DATASET 4,207,344 posts from 567,179 users on Instagram before
We now describe our dataset collection, pre-processing, and Threads launch, and 4,584,317 number of posts from 582,074
statistics in the following sections. users on Instagram after Thread launch.
1) Threads: We use the threads-net API1 to gather Threads For the topical analysis, we exclude reposts to avoid repeat
data. We use the fact that each Threads account is associated counting topics on Threads. This covers 20.83% of Threads’
with an auto-generated numeric ID. Starting from August posts. Note, Instagram does not have a repost feature. Hence,
14th to September 4th, we iterate over all integers from 0 we do not apply this criteria on Instagram. On Threads, 2.88%
to 12 million to identify any account with IDs in this range. posts do not contain any text (i.e. they only contain images),
This approach helps us overcome the limitation of searching preventing us from performing topical analysis. Thus, we also
the Threads platform for specific keywords and provides an exclude these.
1 [Link] 2 [Link]
 (a)Post Counts 1) Overview of Timeline: We start by measuring the number
106 Threads launch Threads of posts per day per user. Figure 1 shows the (a) total post
Instagram
Posts Count

105 counts (solid line) and (b) the average post per user (dashed
line) on two platforms daily from May 5 to September 13. The
104 orange line shows the data for Instagram, while the blue line
(b)Average post per user shows the data from Threads, which is from its launch date.
Threads The grid starts from the first Monday, May 8, in the dataset,
Instagram
Average post

2
per user

and the interval is 7 days.

1 On Threads, the daily average number of posts per user
0 is 1.65 compared to 1.19 on Instagram. We see a significant
2023-05-08
2023-05-15
2023-05-22
2023-05-29
2023-06-05
2023-06-12
2023-06-19
2023-06-26
2023-07-03
2023-07-10
2023-07-17
2023-07-24
2023-07-31
2023-08-07
2023-08-14
2023-08-21
2023-08-28
2023-09-04
2023-09-11
peak in activity on Threads on the second day of its launch
though. To gain deeper insights into the temporal patterns,,
Date we perform auto-correlation analysis. Specifically, we check
if per-day post counts, on each platform, has a correlation with
Fig. 1: (a) Daily volume of posts on Instagram and Threads
the posts counts on the following day.
(b) Daily average of posts per users
Using the daily post count as a time series, we calculate the
autocorrelation function (ACF) separately for each platform.
1.00 1.0
0.75 0.8 Figure 2 shows the autocorrelation function plots of Instagram
Autocorrelation

0.6
Autocorrelation

0.50 and Threads. The x-axis represents the lag in the number of
0.25 0.4
0.00 0.2 days, and the y-axis represents the correlation coefficient. The
0.25 0.0 blue area represents the confidence bands with 95% confidence
0.50 0.2
0.75 intervals. Correlation coefficients outside these bands are sta-
1.00 0.4
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 tistically significant. The ACF plots reveal autocorrelation in
Lag(Day) Lag(Day)
the post count time series data for both platforms. Figure 2(a)
(a) Instagram (b) Threads shows correlation is significant for one (y=0.90) and two
Fig. 2: Autocorrelation for daily post counts: (a) Instagram (y=0.81) days. Albeit, the correlation decreases on the 2nd
(May 4th to Sept 9th.) (b) Threads (July 5th to Sept 9th.) day. Similarly for Threads, Figure 2(b) shows the correlation is
statistical significance of one (y= 0.66) and two days (y=0.46).
Moreover, in the Instagram dataset, 9.38% of images have These findings suggest that activities on both platforms do
embedded text in the images, and these parts of the text have have auto-correlation, and the number of posts on preceding
a median length of 18 words. For these posts, we combine days can predict the number of posts on the following day.
the text in the images (CrowdTanlge provivdes this text) with Moreover, the correlation is higher on Instagram as compared
the users’ added caption. We do so because Instagram is an to Threads, potentially showing that Threads nascent activities
image-centric platform, and if users have to share a text- are still unpredictable.
based message such as a quote, they sometimes share it the 2) Diurnal Post Activity: We next examine the per-hour
image [1]. differences between the two platforms. For this, we separate
Finally, we perform the following text pre-processing on the weekends and weekdays. In addition to the posts counts,
the remaining Threads and Instagram posts’ text: (i) We we look at the accumulated number of users who have posted
only include English language osts, filtered by langdetect during each hour. Past research on other platforms — Twitter,
library.3 41.47% and 41.82% of total posts are in non-English Facebook, and Google+ — indicates that users exhibit varying
languages on Threads and Instagram, respectively. (ii) We activity patterns on different social media platforms [3]. We
remove all hyperlinks. (iii) We remove any mentions for aim to test if post volumes match on both platforms at a given
users (@username). (iv) We remove all non-alphanumeric time during the day.
characters, including emojis, punctuation, and special symbols
(e.g. #, @, $). This leaves 3,533,095 posts from 556,730 users Instagram Before and After the Threads Launch. We
on Threads, 2,471,012 posts from 421,426 users on Instagram partitioned the Instagram datasets (before and after Threads
before Threads launch, and 2,673,157 number of posts from launch) into weekdays and weekends, separately. Then, we
436,223 users on Instagram after Thread launch. split each day into 24 hourly slots and count the number of
posts and users in each hour, across the Instagram datasets
IV. RQ1: T IME - LEVEL ANALYSIS (before and after Threads launch). Figure 3 shows the number
of posts ((a) and (b)) and users ((c) and (d)) as a percentage
Using the above dataset, we first characterise the posting of the total posts and users in any given hour. Note, using
activities on the two platforms. We present an analysis of the percentage gives a normalised representation of the data.
activities divided into weekdays and weekends. The blue lines represent the Instagram data for before Threads
launch period and the orange lines represent the Instagram data
3 [Link] after Threads launch period.
 6 the Threads launch. Figure 4 shows the hourly percentage
5 of posts and active users on Threads and Instagram (after
Post(%)

Threads launch). It is further split in weekdays (Figure 4(a))

4
and weekends (Figure 4(b)). We see that the proportion of
3 Before Threads launch Before Threads launch
After Threads launch After Threads launch posts on both platforms follows a similar trend during most of
0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 the accumulated 24 hour window. The activities drop on both
(a)Hour of day(weekdays) (b)Hour of day(weekends) platforms from midnight to early morning (up to 5:00 AM) and
25 then reach the maximum at 11:00 (Instagram: 22.0%, Threads:
Before Threads launch
Active users(%)

20 After Threads launch 25.0%). Interestingly, the activity levels start to differ from
18:00 hours though. After this time, the overall percentage of
15
posts drops on Instagram, however, on Threads, it achieves
10 Before Threads launch
After Threads launch another peak at around 21:00 with a positive change of 4.3%
0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 from 18:00. On weekends, both platforms follow a largely
(c)Hour of day(weekdays) (d)Hour of day(weekends) similar pattern of active users and reach a peak at 11:00
Fig. 3: Compared Instagram before and after Threads launch: hours with 15.7% and 15.0% for both Instagram and Threads,
(a) and (b) show the percentage of accumulated posts in each respectively.
hour on weekdays and weekends, respectively. (c) and (d) Moreover, for the hours with similar trends on both plat-
show the percentage of accumulated active users in each hour forms, we observe a lag between the two platforms, with Insta-
on weekdays and weekends, respectively. Times are converted gram leading Threads as shown in (Figure 4(a)) and weekends
to EDT zone. (Figure 4(b)). To understand this time lag, we perform a cross-
correlation function (CCF) on the percentage of activities. We
First, we look at the posts in two periods for weekdays take each data series (Instagram and Threads) and use the CCF
(Figure 3(a)) and weekends (Figure 3(b)). We observe that to identify lags of one series that might be useful predictors
the accumulated percentage of posts on the two platforms of another. We take the activities on Instagram as xt and
have similar patterns on weekends and weekdays. We note Threads as yt , separately on weekdays and weekends. The
that the posts volume on the platform starts to increase from results show that the most dominant cross-correlations occur at
early morning (05:00) and reaches its peak at 11:00. It then a lag of -1 hour both on weekdays (cross correlation=0.919)
begins a downward trend with marginally stable activities from and weekends (cross correlation=0.961). The results of the
16:00 until 20:00. Similar patterns hold for both weekends and cross-correlation analysis fall outside the bounds of the 95%
weekdays. confidence interval (from -0.4 to 0.4). This indicates statistical
significance. The results mean that on the aggregate post
However, in terms of users, we see contrasting results,
activities time series the activities on Instagram lead the
especially on weekdays. Figure 3(c) shows that the overall
activities on Threads by 1 hour.
trends of active users in hourly windows are similar on
We then compare the count of active users in each hour.
Instagram (both after and before the Threads launch). The
Figure 4(c) and Figure 4(d) plot the active user counts accu-
number of active users peaks between 11:00 – 12:00 AM.
mulated per hour on weekdays and weekends separately on
Interestingly, the proportion of active users has dropped after
Instagram and Threads. The blue dashed-line is higher than
the Threads launch with a difference of 2.5% at the peak hour.
orange dashed-line in Figure 4(c). This means there are more
On weekdays, overall proportion of active Instagram users
users active on Threads at any given hour during weekdays as
is 17.7% before the Threads launch, and 15.6% on after. In
compared to Instagram. Users are counted for each hour they
contrast, the proportion of active users does not change on
post. We further measure the average number of different hours
weekends (average of 11.0%) before and after, as shown in
where users posts on each platform. On weekdays, the average
Figure 3(d).
number of different hours a user posts is 4.65 and 3.74 for
For Instagram, we find that during the weekends there is Threads and Instagram, respectively. Given that we have more
no difference in terms of active users and the posts count users on Threads, as compared to Instagram, after the launch
in an hourly proportion before and after the Threads launch. and a higher average number of hours per user, the higher
However, there are differences in terms of user counts. We number of active users is justified. Compared to Instagram,
find that the proportion of active users during weekdays has a greater percentage of users are active on Threads during
reduced after the Threads launch. However, there is no change the night from 21:00 to 4:00 on both weekdays (38.47%) and
in terms of the post pattern on the platform. Overall, we weekends (11.73%). The biggest difference in hours is 22:00
observe a reduction in the proportion of users during the on weekdays and 0:00 on weekends. Threads has a higher
weekdays, but there is no change in the proportion of post percentage of active users at night than Instagram.
counts in hours. Overall, we find that within Instagram, the proportion of
Comparison between Instagram and Threads. Next, we post distribution in the hours window does not differ before
compare the hourly activity patterns on Threads with Insta- and after the Threads launch. However, the proportion of active
gram. For this, we only consider the Instagram data after users in hours changes before and after the Threads launch.
 6 BERTopic [14], which employs sentence transformers and c-
5 TF-IDF to generate topics. We combine the posts from both
Post(%)

platforms to create one corpus to ensure consistency and

4
coherence in topic comparisons [4]. Using a threshold of at
3 Threads Threads
Instagram Instagram least 3,500 posts per topic, we identify 206 topics across the
0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 combined corpus. These topics cover 50.27% of posts in the
(a)Hour of day(weekdays) (b)Hour of day(weekends) data.
25 Threads Topic Volume. We first look at the most common topics
Active users(%)

20 Instagram
within each platform. We see that the prominence of topics
15 varies across the two platforms. Figure 5 presents the absolute
10 Threads number of posts (bars) and the normalized number of posts
Instagram per user (red line) for the top 30 topics. The x-axis shows
0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 the topics ranked in order of their post count in the Instagram
(c)Hour of day(weekdays) (d)Hour of day(weekends)
dataset before the Threads launch. Normalization is performed
Fig. 4: Compared Threads and Instagram after Threads launch: by calculating the number of users engaging in a given topic.
(a) and (b) show the percentage of accumulated posts in each Figure 5(a) and (b) have similar trends in terms of both the
hour on weekdays and weekends, respectively. (c) and (d) total number of topics and the normalized result. Some topics
show the accumulated active users in each hour on weekdays vary on Instagram before and after Threads launch in the total
and weekends, respectively. Times are converted to EDT time number of posts. These topics are related to specific events
zone. occurring during that period, such as mother, barbie
(film) and lgbtq, which refer to Mother’s Day, film-
(a)Instagram data before Threads launch.
100000 4 release and LGBT pride, respectively.
Normalized Post Normalized Post Normalized Post
Post Counts

(by User)

50000 2 Figure 5(c) presents the topics on Threads. When compared

with Figure 5(b), we notice that topics about art (music
0 0 and art), food (food, coffee and wine), and sports
(b)Instagram data after Threads launch.
100000 4 (football and gym) are popular on both Instagram and
Post Counts

(by User)

Threads. However, Threads also has distinct topics from

50000 2
Instagram. Topics on social media applications (Threads and
0 0 Twitter) are more discussed on Threads. Politics(trump)
(c)Threads is another topic discussed more on Threads than on Instagram.
100000 4
Post Counts

However, discussion patterns on the two platforms display a

(by User)

50000 2 different trend. A total of 24,659 (4.43%) users wrote about

topic trump on Threads. Yet there are only 9,113 (2.09%)
0 0
users writing about trump on Instagram. Interestingly, each
music
art
birthday
food
fashion
mother
football
gym
flowers
school
book
lgbtq
wedding
wine
dog
photo
home
god
trump
hair
car
friday

iphone
barbie
twitter
coffee
business
life
rain

threads

user on Instagram is more active though, with 2.26 posts per

Topics
user about trump vs. 1.43 posts per user on Threads. Fur-
thermore, Threads has fewer posts on daily lifestyle topics like
Fig. 5: Posts count for the top-30 topics (left-y) and normalised fashion (fashion and hair) and pets (dog) as compared
posts per user(right-y). (a) Instagram before Threads launch. to Instagram, where these topics are popular [15]. Moreover,
(b) Instagram after Threads launch. (c) Threads. The x-axis users prefer to share festivals and anniversary topics such as
is ranked based on the number of posts on Instagram before birthday and wedding on Instagram. Instagram has a
Threads launch. higher total number of posts on these topics than Threads.
3,850 users post wedding relevant content on Threads, and
When being compared with Threads, the Instagram volume
12,855 users post wedding on Instagram. 925 (7.20% of
of posts on Instagram leads Threads by one hour. However,
12,855) users post this topic on two platforms. 11,111 users
the overall pattern of content generation and users being active
post birthday on Threads, and 44,677 users on Instagram.
remains similar on both platforms, with the maximum number
There are 1,914 (4.28% of 44,677) common users who post
of users being active around noon, and hence the most posts
this topic on two platforms. Finally, we also notice a topic AI
are also shared at the same time.
outside Figure 5. 10,509 (1.89%) users engaging in discussion
V. RQ2: C ONTENT- LEVEL ANALYSIS on ai with an average of 2.12 posts per user on Threads. Only
We now analyse the topics on both platforms. We also look 3,282 (0.75%) users discussing ai on Instagram. It indicates
at the temporal evolution of topics to understand whether users a potential interest in AI-relevant topics on Threads that might
continue a similar discourse on both platforms. not be as prominent on Instagram.
1) Topical Discussions: We first look at the most common Overall, we note that while there are topical similarities
topics on each platform, and their weekly trends. We use on the two platforms, users on Threads post less on fashion-
related topics and more on topics related to politics and 3) Feedback: We also measure feedback on posts, using the
technology than Instagram. number of likes for each posts on both platforms. To calculate
Topics Over Time. We then check the temporal variations of the feedback of a post, we first calculate the mean value of
the top 20 topics on two platforms. To do so, we compare the likes of all posts published before this post. Taking the
the weekly volume of each topic over the duration of dataset. mean value of the previous likes and comparing it with the
Figure 6 shows the normalised number of posts for a given likes received by the current post helps in measuring whether
topic over time. To be more precise, for each topic we calculate the new post received better or worse engagement. We use the
the total number of posts across all the dataset, and divide obtained mean value to divide the likes count by the current
the weekly count by the total number. Hence, giving us a post, and use Laplace smoothing to avoid a denominator of
normalised change over time, and each row will sum up to zero.
100 percent. The x-axis ticks show all the Sundays. Sunday B. Intra-platform exploration
here is the end of the week. Figure 6(a), (b) and (c) share the
same colour scale. On Threads, we notice that the discussion We now estimate this feedback measure’s impact on topic
about threads and twitter on Threads mainly stem from consistency at post and user levels. We hypothesize that topic-
the first week after Threads launch. It also spurs discussion consistent posts tend to have worse feedback since users with
on these two topics on Instagram in the same week. However, unpopular posts tend to focus more on their interests (narrow
combining the results from the previous section, the volume interests) compared with more popular users [16].
of this discussion is relatively small compared to Threads. On For this, we first divide all the posts into topic-consistent
both platforms, the discourse on threads and twitter posts and topic-inconsistent posts. From all the posts that
rapidly dissipates. Most topics on Threads have a decreasing have been assigned a valid topic from BERTopic model (same
trend, except trump. The number of posts on trump keeps model from RQ2), there are 709,272 (28.25%) topic-consistent
a stable trend, slightly increasing in two months. It hints at a posts and 1,801,559 (71.75%) topic-inconsistent posts on
potential interest in political discussion on Threads. Instagram, while 115482 (8.93%) topic-consistent posts and
1178260 (91.07%) topic-inconsistent posts on Threads. There
VI. RQ3: T OPIC - CONSISTENCY EXPLORATION are fewer consistent topics posts on Threads than on Instagram
Finally, we analyze the impact of the feedback received by as a percentage.
posts in individual users’ topical choices. We explore whether We now analyse the feedback differences on topic-consistent
users’ topical choices change as they move from Instagram to and inconsistent posts. Figure 7 shows the distribution of
Threads. As such, we first define the term topic consistency feedback for both Threads and Instagram. On Threads, the
and then explore the changes in users’ topic consistency within first quartile of posts have zero feedback because some posts
and across the platform. have no likes on the newly launched Threads. We find that
the feedback value on Instagram (median = 0.95, std =
A. Topic-consistency and feedback 10853.97) is higher than that on Threads (median = 0.55,
We define topic consistency as continuing the same topic std = 180.74) with the K-W test (H(1) = 33302.96, p-value
in two consecutive posts. We further split this into two levels <0.001). In general, topic-inconsistent posts (median = 0.88,
to examine topic consistency at the intra- and inter- platforms std = 9637.13) have relatively higher feedback than topic-
scale. consistent posts (median = 0.79, std = 7800.72) (with K-W
1) Intra-platform topic-consistency: If two consecutive test results: H(1) = 770119.22, p-value <0.001 on Instagram
posts by a user on a single platform share the same topic, the and H(1) = 62099.37, p-value <0.001 on Threads).
preceding post in time will be labeled as having Intra-platform Unsurprisingly, these results indicate that topic-inconsistent
topic consistency. posts are more likely to have better feedback, i.e. users who
2) Inter-platform topic-consistency: Inter-platform topic talk about a topic that receives a high number of likes are more
consistency measures a user’s continuation of the same topic likely to change the topic of their next post. This phenomenon
from one Instagram to Threads. is more common on Instagram. A potential explanation for
To explore the topic consistency of users when migrating this phenomenon is that users who prefer to publish posts on
from Instagram to Threads, we inspect posts for each user. We a wide variety of topics tend to be popular with audiences.
first extract the topics of n recent posts on Instagram before Naturally, receiving positive feedback reinforces this message
the launch of Threads (the specific number of posts will be for publishers. But users who focus on same topics care little
discussed as a threshold in Section VI-C). We then extract the about the feedback.
topics of n posts on Threads. If there is an overlap between
these two sets, the user will be regarded as having inter- C. Inter-platform exploration
platform topic consistency. However, it is difficult to select Next, we utilize the inter-platform consistency definition
the optimal threshold (n) for the number of posts to calculate (users writing similar topics on Instagram and Threads in their
this overlap. Hence, we experiment with three thresholds (i.e. consecutive posts) and divide users into consistent and non-
, the first or latest 1/2/3 posts to calculate the overlap) and consistent groups. We test three different thresholds on number
compare the results of these three alternatives. of consecutive posts (from 1 to 3) to define the inter-platform
 music 4 8 9 9 9 9 9 19 20 5 5 7 7 7 8 8 8 9 16 16 7 24 20 11 9 8 8 7 5 5 4 0
art 4 9 9 9 9 9 9 17 18 5 6 8 8 8 8 8 8 9 15 16 7 24 20 10 8 8 7 7 5 5 5 0
food 5 9 10 11 10 10 10 15 16 5 6 8 9 9 9 9 8 10 13 13 6 29 24 11 8 7 6 6 4 3 3 0
football 3 7 6 6 6 6 6 27 27 6 2 4 4 4 4 5 5 10 24 27 10 24 17 10 8 8 8 8 5 6 7 0
birthday 5 10 10 10 11 11 10 14 14 4 6 9 9 9 9 9 9 10 12 12 5 23 22 11 8 9 7 7 4 4 4 0
gym 3 9 9 9 9 9 9 18 19 5 6 8 8 8 8 8 8 9 15 15 7 23 26 12 8 7 7 6 4 3 3 0
fashion 4 9 9 9 9 8 9 18 19 5 5 7 7 7 7 7 7 10 16 18 9 25 23 11 8 7 7 6 4 4 4 0
trump 2 6 5 4 4 5 5 29 33 7 2 3 4 4 5 5 4 12 25 23 13 11 11 8 7 10 8 10 10 10 16 0
god 4 7 8 8 8 9 8 20 21 5 5 7 7 7 7 7 7 10 17 17 8 18 20 11 9 8 9 8 6 6 6 0
book 3 7 7 7 8 8 8 21 23 6 5 7 6 6 7 7 6 9 19 20 9 23 19 10 8 8 7 7 5 6 6 0
threads 4 5 6 8 5 7 7 22 24 10 58 11 4 3 2 2 2 4 5 7 3 70 13 5 3 2 2 2 1 1 1 0
school 4 10 13 21 11 13 10 10 7 1 2 3 4 3 7 14 17 14 16 19 3 20 17 9 7 8 9 11 7 6 6 0
friday 6 10 11 11 10 10 10 15 15 4 7 9 9 8 8 8 9 10 14 14 5 29 25 11 7 6 6 5 3 3 3 0
dog 5 10 9 9 9 9 9 17 17 6 5 7 7 7 7 7 7 20 14 13 6 32 21 10 7 6 5 5 5 3 4 0
wine 5 9 10 10 10 10 9 16 16 4 6 8 9 9 9 8 8 9 14 14 5 31 23 10 8 7 6 6 3 3 3 0
photo 4 10 10 10 11 10 10 15 16 4 6 9 9 8 9 8 10 9 13 13 6 35 21 10 7 6 5 6 4 3 3 0
flowers 5 12 11 12 11 10 10 13 12 3 7 9 9 9 9 9 9 9 12 12 5 26 22 11 8 7 6 7 4 4 4 0
business 3 8 8 7 7 8 9 22 23 6 5 7 7 7 6 6 7 9 19 18 9 24 23 12 9 7 6 6 4 4 4 0
rain 4 9 8 7 8 9 11 19 19 6 5 9 8 8 7 7 11 10 17 13 4 22 24 11 9 5 5 11 5 4 4 0
home 4 10 11 10 10 11 10 15 16 3 5 9 9 9 9 9 9 10 13 12 7 26 22 11 8 7 6 6 4 4 5 0
twitter 2 5 3 5 5 6 5 25 35 9 28 6 5 9 5 3 2 6 16 14 6 77 11 3 3 2 1 1 1 0 0 0
05-07
05-14
05-21
05-28
06-04
06-11
06-18
06-25
07-02
07-09
07-09
07-16
07-23
07-30
08-06
08-13
08-20
08-27
09-03
09-10
09-17
07-09
07-16
07-23
07-30
08-06
08-13
08-20
08-27
09-03
09-10
09-17
(a)Instagram (b)Instagram (c)Threads
before Threads launch after Threads launch
Fig. 6: The 3 heatmaps for top 20 topics on (a)Instagram before Threads launch, (b)Instagram after Threads launch and
(c)Threads. Normalise by the number of each topic/total number on this platform.

Topic-consistent posts Topic-inconsistent posts 34.04%/24.03% on Threads when threshold = 1).

4 These results indicate that users who maintain topic con-
sistency when they migrate from the “parent” platform to the
Feedback Value

3
“child” are also more likely to publish posts on the same topics
2
consecutively. Further, the same group of users is more likely
1 to maintain topic consistency on Instagram but change topics
0 more frequently on Threads.
Instagram Threads We finally analyze the differences in feedback on posts
Platforms
from these user groups and show the distribution of feedback
Fig. 7: Box plot of feedback value for both topic-consistent in Figure 9(a) and Figure 9(b) for Instagram and Threads,
posts and inconsistent posts on Instagram and Threads. respectively. The median feedback count on Threads is less
than 1 for both consistent and inconsistent users on all
Threshold =1 Threshold =2 Threshold =3 thresholds. On Instagram, the median value is close to 1 for
Consistent users 46,051 84,631 102,807 both user types. A K-W test shows that there is no significant
Inconsistent users 131,293 92,713 74,537 difference in feedback received by both user-groups, even
TABLE I: The number of consistent/inconsistent users for though users’ feedback on Instagram (median = 0.98, std =
different threshold. 15528.53) is relatively higher than that on Threads (median =
0.67, std = 118.56). These results indicate that there is little
topic consistency. A threshold of 3 posts means that we require correlation between whether users maintain topic-consistency
users to repeat a topic at least once in the 3 consecutive at inter-platforms scale and the feedback they receive. And
posts to be considered topic-consistent. A higher threshold generally users receive better feedback on Instagram compared
will increase the number of topic-consistent users. Conversely, with Threads.
a lower threshold (1 post) will lower the number of topic-
consistent users. Table I shows the number of consistent and VII. C ONCLUSION
inconsistent users with all thresholds. Our study serves as the first characterization of Threads.
We first analyze the percentage of topic-consistent posts We focus on the inteheritted overlaps between Threads (the
by topic-consistent and -inconsistent users. Figure 8 shows “child”) and Instagram (its “parent”). We have inspected the
these percentages for both Instagram and Threads with all posting activities and topical discourse of the early adopters of
thresholds. This figure shows that the percentage of topic- Threads. We have also compared the same users on Instagram
consistent posts (with threshold=1), on Instagram, from consis- to record the difference in users’ communication on both
tent users (40.83%) is higher than inconsistent users (34.14%). platforms.
The corresponding percentage on Instagram is also higher than Our analysis (RQ 1) has shown that both Instagram and
Threads (i.e. 40.83%/34.14% on Instagram is higher than Threads follow a similar posting pattern across 24-hour accu-
 Instagram-consistent users Threads-consistent users R EFERENCES
Instagram-inconsistent users Threads-inconsistent users
40 [1] L. Manikonda, V. V. Meduri, and S. Kambhampati,

Percentage of topic-consistent posts

35
30
“Tweeting the mind and instagramming the heart: Ex-
25 ploring differentiated content sharing on social media,”
20
15
in AAAI ICWSM, 2016.
10 [2] C. Zhong, H.-w. Chang, D. Karamshuk, D. Lee, and
5 N. Sastry, “Wearing many (social) hats: How different
0
Threshold = 1 Threshold = 2 Threshold = 3 are your different social network personae?” in AAAI
Thresholds
ICWSM, 2017.
Fig. 8: Percentage of topic consistency posts for consis- [3] B. H. Lim, D. Lu, T. Chen, and M.-Y. Kan, “# mytweet
tent/inconsistent users on two platforms (with three different via instagram: Exploring user behaviour across multiple
thresholds). social networks,” in IEEE/ACM ASONAM, 2015.
[4] J. He, H. B. Zia, I. Castro, A. Raman, N. Sastry, and
Consistent Users Inconsistent Users Consistent Users Inconsistent Users G. Tyson, “Flocking to mastodon: Tracking the great
4 3.5 twitter migration,” in ACM IMC, 2023.
3.0
[5] H. Kwak, C. Lee, H. Park, and S. Moon, “What is twitter,
Feedback value

Feedback value

3 2.5
2.0 a social network or a news media?” in ACM WWW, 2010.
2 1.5
1.0
[6] Y. Li, Z. Zhang, Y. Peng, H. Yin, and Q. Xu, “Matching
1
0.5 user accounts based on user generated content across
0 0.0 social networks,” Future Generation Computer Systems,
threshold = 1 threshold = 2 threshold = 3 threshold = 1 threshold = 2 threshold = 3
Posts of users on Instagram Posts of users on Threads 2018.
(a) Feedback (Ins) (b) Feedback (Threads) [7] M. Han Veiga and C. Eickhoff, “A cross-platform col-
lection of social network profiles,” in ACM SIGIR, 2016.
Fig. 9: (a) Box plot of feedback for consistent/inconsistent
[8] R. Farahbakhsh, A. Cuevas, and N. Crespi, “Character-
users on Instagram. (b) Box plot of feedback for consis-
ization of cross-posting activity for professional users
tent/inconsistent users on Threads.
across facebook, twitter and google+,” Social Network
mulated cycles. Both platforms have peak activities at noon. Analysis and Mining, vol. 6, pp. 1–14, 2016.
However, the proportion of active users on Instagram is lower [9] S. Kumar, R. Zafarani, and H. Liu, “Understanding user
after the Threads launch, albeit the proportion of posts in the migration patterns in social media,” in AAAI AAAI, 2011.
hour window remains consistent. In addition, we note that [10] C. Fiesler and B. Dym, “Moving across lands: Online
Instagram leads the activities on Threads by 1 hour. We have platform migration in fandom communities,” ACM on
also shown that there is a difference in topical focus across HCI, no. CSCW1, 2020.
the two platforms (RQ 2). Instagram has a higher percent- [11] U. Jeong, P. Sheth, A. Tahir, F. Alatawi, H. R. Bernard,
age of topics related to lifestyle and fashion, while Threads and H. Liu, “Exploring platform migration patterns be-
has more topics related to politics and technology. Finally, tween twitter and mastodon: A user behavior study,”
we have explored the topic-consistency across platforms and arXiv preprint arXiv:2305.09196, 2023.
compare the feedback (RQ 3). We find that users care more [12] E. Newell, D. Jurgens, H. Saleem, H. Vala, J. Sassine,
about the content itself than about the feedback their posts C. Armstrong, and D. Ruths, “User migration in online
receive, especially on Instagram. We also notice that users who social networks: A case study on reddit during a period
maintain topic consistency when they migrate from Instagram of community unrest,” in AAAI ICWSM, 2016.
to Threads are more likely to publish posts consecutively on [13] U. Jeong, A. Nirmal, K. Jha, X. Tang, H. R. Bernard,
the same topics. and H. Liu, “User migration across multiple social media
platforms,” arXiv preprint arXiv:2309.12613, 2023.
Limitation and Future Work. This is the first work on [14] M. Grootendorst, “Bertopic: Neural topic modeling
Threads. However, we wish to flag several limitations mainly with a class-based tf-idf procedure,” arXiv preprint
arising from the data collection. First, we are limited by arXiv:2203.05794, 2022.
25 posts per user on Threads. Second, some Threads users [15] Y. Hu, L. Manikonda, and S. Kambhampati, “What we
cannot be associated with their Instagram accounts due to their instagram: A first analysis of instagram photo content
privacy settings or username changes. Third, our longitudinal and user types,” in AAAI ICWSM, 2014.
analysis is limited to a two month duration and cannot [16] E. Ferrara, R. Interdonato, and A. Tagarelli, “Online
necessarily be generalized to the whole platform. We plan popularity and topical interests through the lens of in-
to extend our data collection and the observation period to stagram,” in ACM HT, 2014.
expand our findings, and holistically characterize the users’
movement from a parent to child platform. Other potential
research areas include the impact on the parent’s platform
activities, e.g. related to Instagram’s e-commerce features.