Are likes, shares, or comments more important for TikTok video success?
Summary
TikTok’s recommendation algorithm is an iterative process
which repeatedly displays videos to larger and larger test audiences, until
engagement from the test audience falls below some threshold. It is unknown how
“engagement” is measured, but it is usually speculated to be some combination
of likes, comments, and shares. Various people have hypothesized the weights of
these metrics (e.g. one share = 10 likes), but, to our knowledge, no one has
ever measured it. We therefore attempted to measure this on a representative
sample of ~20,000 videos.
Our research concludes that these metrics are highly
collinear and no predictive value is found from comment and share count. (I.e.
you should just optimize for likes, and don’t need to optimize for shares or
comments.) You will need about 1/5 of your viewers to like your video in order for it to be promoted.
Background and Motivation
More background on the structure of the TikTok algorithm can
be found here.
Veed
contains one of the more technical explanations of the TikTok algorithm. They hypothesize
the following:
Now that TikTok knows some basics about the video, it will
now boost the video to a small number of users. After booting [sic] the video, an
evaluation will happen based on how the sample set of user interaction with the
piece of content. Each metric that is tracked has an associated score that
varies in value. Here is an estimated score hierarchy that awards content per
user interaction.
Rewatch rate = 10 Points
Completion rate = 8 Points
Shares = 6 Points
Comments = 4 Points
Likes = 2 Points
Our understanding is that this score hierarchy was based on
the author’s intuitions, as opposed to any evidence, and we are not aware of
any other research into possible score hierarchies. Nonetheless, many creators
include comment or sharebased call to actions (e.g. “share with your tallest
friend”). Such CTA’s are commonly recommended for other platforms, e.g. YouTube.
It’s therefore important for creators to understand which
metrics influence recommendations, so that they can incorporate CTAs
appropriately.
Methodology
A representative sample of ~20,000 videos was used and
partitioned into ~12,000 videos from 2020 and ~5,000 videos from 2019. Each
video was tagged with the number of likes, views, shares, and comments it received.
LASSO, Ridge, and ordinary least squares regressions were run against each data
set, with both linear and logtransformed values. LASSO and Ridge hyperparameters were
selected via fivefold cross validation.
Results
Year

Transformation

Regression Type

Shares Coefficient

Comments

Likes

Intercept

Alpha

R^2

2020

None

Lasso

0

0

4.556428

10416.2

3.72E+07

0.798259

2020

None

Ridge

25.16904

0.26108

4.410611

9774.501

1.00E+08

0.806881

2020

None

OLS

27.0471

0.25979

4.396494

9752.731

N/A

0.838235

2020

Logarithm

Lasso

0.03346

0.05088

1.078338

1.641511

0.012661

0.945464

2020

Logarithm

Ridge

0.04422

0.07564

1.099521

1.606834

1

0.945643

2020

Logarithm

OLS

0.04424

0.07572

1.099592

1.606708

N/A

0.948446

2019

None

Lasso

0

0

6.705435

3381.721

1.10E+06

0.837942

2019

None

Ridge

2.62557

8.774306

6.710484

3340.445

1.00E+08

0.838817

2019

None

OLS

6.47772

52.20791

6.501989

3196.172

N/A

0.862279

2019

Logarithm

Lasso

0

0

1.037677

2.036934

0.021131

0.889898

2019

Logarithm

Ridge

0.013935

0.01206

1.039334

2.031968

5

0.890604

2019

Logarithm

OLS

0.013507

0.01301

1.040213

2.030331

N/A

0.901423

2020 Regular
The following graphs compare the actual number of views a video got to the number of views the labeled regression predicted them to receive.
2020 Logarithm
2019 Regular
2019 Logarithm
Discussion
The fit is reasonably good, as demonstrated both by the
formal R2 near 0.9 and eyeballing the plots.
LASSO regression drops share and comment count variables
from the untransformed data, except for the log transformed 2020 data, where
they are assigned small values. There are several reasons to think that comment
and share count adds little predictive value:
 First, simply based on priors, we might assume that the independent variables are all colinear
 Testing confirms that this is, indeed, the case
 OLS results seem rather implausible (e.g. one comment = 17 likes)
 Dropping these variables causes only minor degradations to predictive fit
Of course, simply because these variables are not predictive
within our sample, that may not mean that they have no predictive value. It
could be that so few creators use comment or sharebased CTA’s that comment,
share, and like counts are heavily correlated, but we would see different
results if more creators use these CTAs. A larger sample size may fix this.
Conclusion and Recommendations
It seems likely on both priors and evidence that engagement metrics
are correlated, and that optimizing for one over the others is unwarranted.
Creators should simply try to create content people like –
comments and shares will follow naturally.
This research done in collaboration with @lilweehag. The code underlying this post can be found here.
Interestingly enough, the summary is essentially the principle behind inbound marketing.
ReplyDeletePlzz like comment share plzzz plzzzzzzz
ReplyDeleteNice to be visiting your blog again, it has been months for me. Well this article that i’ve been waited for so long. I need this article to complete my assignment in the college, and it has same topic with your article. Thanks, great share. Tik tok counter
ReplyDeleteI'd like to express my gratitude for writing such an informative piece on this blog. This article provided me with a variety of data buy tiktok likes. Continue to post.
ReplyDeleteVery efficiently written information. It will be beneficial to anybody who utilizes it, including me. Keep up the good work. For sure i will check out more posts. This site seems to get a good amount of visitors. Tiktok counter
ReplyDelete