Designing a Content Hotness Scoring Algorithm for Community Platforms

In the process of a community redesign, the big‑data team combined business details, research, discussion, and trial‑and‑error to design a basic content hotness scoring algorithm.

Reference: A few foreign companies (Hacker News, Reddit, Stack Overflow, StumbleUpon) have publicly described their hotness algorithms; links are provided for further reading.

Data dimensions considered include:

Time dimensions: post_time, last_reply_time, last_op_time.

Interaction dimensions: view_num, reply_num, favor_num, like_num, reply_like_num, share_num.

Content dimensions: content_length, reply_avg_length, picture_num.

User dimensions: user interest, activity, reputation, etc.

The business meaning of each dimension is explained: time controls decay, interaction metrics are weighted to reflect their impact, content metrics serve as auxiliary quality signals, and user metrics enable personalized recommendation.

Algorithm design focuses on the last_reply_time as the time dimension and models decay using Newton’s cooling law:

H(t) = H_a * exp[-γ * (t - t_last) / 86400]

The raw hotness value H_a is calculated from interaction metrics:

H_a = ln(1 + N_view) + 1.0 * N_reply + 1.75 * N_like + 3.2 * N_favor

Reading count is transformed with a natural logarithm to reflect diminishing contribution, and content length contributes via lg(N_length).

The final hotness formula combines all dimensions:

H = [lg(N_length) + ln(1+N_view) + 1.0*N_reply + 1.75*N_like + 3.2*N_favor] * exp[-γ*(t - t_last)/86400]

User interest vectors are built from recent behavior, normalized to [0.1, 1.0], and combined with content tag vectors; the total score may be further adjusted by a random factor to diversify ordering:

rand(0.75,1.25) * C_interest * H

Conclusion: The algorithm balances technical decay, weighted interaction, content characteristics, and user interest to produce a dynamic hotness score that reflects both content quality and personalized relevance.