New Trending Algorithm - Attempt Number 5

A brief note in advance: For those who have not yet noticed: I have been among the Top20 witnesses of this blockchain for a few days now. Many thanks to Steemchiller and to the many voters who believe in me. I hope we can breathe more life back into Steem together. So, now back to work and to the actual topic.

My last attempt at establishing a new trending algorithm ran for some time on my test domain (https://steemit.steemnet.org/trending).
However, I had to realise that this was not yet the big hit, as the posts that appeared there were not really of any particular significance.

New approach

So a new approach was needed. To do this, I asked myself the question: What qualities must a post have to be 'trending' or 'hot'?

• Payout level (but with less influence than before),
• Number of resteems (higher = better),
• Number of comments (higher = better),
• Number of commenters (higher = better),

I asked DeepSeek about this and received the following suggestions:

1. resteems are typically given the highest weight as they signal strong agreement
2. unique commenters are weighted higher than comment counts alone
3. logarithmic scaling prevents individual metrics from dominating
4. combination of old and new system (retains the proven base score, but supplements it with interaction signals).
5. temporal decay function (older content receives a decreasing score).

DeepSeek had also provided an implementation, but it was unusable because it did not respect Hivemind's database structure at all. Although I had sent DeepSeek the link to the repo, it could not produce any usable results.

From this, I have created a few premises for the new algorithm:

• For reasons of efficiency, there should only be a few additional database queries. Unfortunately, the database structure of Hivemind is such that it is not possible to query all comments on a post with reasonable effort. This requires a recursive call. I currently want to avoid this. Proposal no. 2 is therefore not applicable.
• The order of magnitude of the old score should not be changed significantly so that no replay of the database is necessary. I first have to find an appropriate calculation. I also had to analyse the previous algorithm.
• Older posts should continue to fade into the background after a while. This also required an analysis of the previous algorithm, as this already causes older posts to slide to the back.

Analysis of the current calculation

Up to now, the (official) score has been calculated using the following function:

def _score(rshares, created_timestamp, timescale=480000):
    """Calculate trending/hot score.

    Source: calculate_score - https://github.com/steemit/steem/blob/8cd5f688d75092298bcffaa48a543ed9b01447a6/libraries/plugins/tags/tags_plugin.cpp#L239
    """
    mod_score = rshares / 10000000.0
    order = math.log10(max((abs(mod_score), 1)))
    sign = 1 if mod_score > 0 else -1
    return sign * order + created_timestamp / timescale

The calculation corresponds to that from the steemd tags plugin.

The lines

mod_score = rshares / 10000000.0
order = math.log10(max((abs(mod_score), 1)))

have the effect that rshares <= 1e7 are not taken into account (mod_score <= 1 and log10(1)=0) and high rshares no longer lead to a significant increase in the score.
Finally, the absolute value order is assigned with the sign of the votes. This value is then added to the result from the division of timestamp and timescale.
For a long time, I didn't realise how the division worked and what it meant. It was only when I thought about suggestion no. 5 for a while that I realised this. At this point, I would like to mention that I also asked DeepSeek about this. He'd been throwing wild differences between timescale and decay at me that made no sense from my point of view.

This is because created_timestamp / timescale already cause a decay function. As timestamp increases continuously with every second, the final score also increases. The value for timescale causes the final score to change per second simply by the passage of time. Sounds simple, doesn't it? For DeepSeek, it was probably too simple.
For the new algorithm, this means that the quality component of the score must be dimensioned in such a way that the passage of time still has an effect.

New implementation

Based on the above premises, I have now implemented the following solution:

• There is a combination with the previous score and the new interactions score.
• The interactions score takes into account the number of reblogs and the number of comments.
• Votes weight = 0.2 (only very little influence),
• Interactions weight = 0.9 (higher influence),
• Reblog weight = 1.0 (higher influence compared to comments),
• Comment weight = 0.7.

Several days ago, I used an early version of the new approach to record the essential data for 5 days so that I could easily understand the effect of further adjustments. Here you can see the diagram of the rank changes for selected posts:

You can see very clearly that the inclusion of reblogs and comments alone would have led to numerous ranking changes over time.
About a week ago, I put the new algorithm online at https://steemit.steemnet.org/trending. It took a few days for the new posts to overtake the posts with the old score, but now you can see - I think - much better results in the trending posts (but it's a coincidence that a post about me was created just today and - obviously because of the many congratulators - it lands quite high up).

What do you think? Do you have any suggestions for improvement?

Let me know what you think about the new attempt.

08.06.2025

Steem Search on https://moecki.online/ My Full Node on https://api.moecki.online For fast account recovery set your recovery account to @moecki.recovery. My Github Vote here