RISE-heat-baseline

A small, reproducible research baseline trained on real, authorized aggregate X post observations. It did not beat a simple baseline in the measured comparison.

源码：RISE / x-quality-growth。算法研究：固定版本公式 · 时间参数。这是独立研究实验，并非 X 官方模型。

What was actually trained

PyTorch 2.14.0, CPU, one thread, 400 Adam steps per fit. Each post has a separate curve:

count(age) = b + A * (1 - exp(-age / tau_seconds))

Three positive float64 trainable parameters per curve, enforced through softplus. The saved release contains two fitted curves, six learned scalar parameters in total. The JSON state is 2,929 bytes; the raw parameter payload is 48 bytes across the two fits. Training uses mean squared log1p-count error. b is a fitted intercept, not an invented age-zero observation.

The two final fits use 19 observations across two posts, which are correlated observations, not 19 independent examples. Seven evaluable PyTorch rolling origins all belong to one launch post. A fitted curve's time constant is not an X algorithm parameter or proof of a post's lifespan.

Honest backtest

Only information available before each target observation enters that fit. Public views and owner analytics impressions remain separate. No random row-level train/test split.

Latest held-out launch capture:

On the same five 15–60 minute origins, all from the same post, MAE was 9.40 / 12.30 / 28.30, respectively. The fitted model is worse on this comparison. These are descriptive retrospective errors, not evidence of generalization, significance, causal growth, or reliable future viral predictions. Complete origins and errors are in backtest.json.

Inputs, outputs and boundaries

Current inputs: real publication age and previous cumulative exposure counts. Current output: an exploratory within-post future count. Text, images, video, attachments and comments have not been trained as features. A general input contract is documented in model inputs, with future frozen text/vision features and cutoff-aware comment statistics. Other platforms have not been calibrated.

Public views may include operator opens and repeated impressions. Owner analytics lacks enough repeated captures for training at this release. No unique-user reach, negative-feedback probability, or attributed new followers is fabricated. Forecast intervals and parameter intervals are unknown. The official 8-hour half-life popular-pool score is a separate source heuristic, not a calibrated 24-hour prediction.

Use

Download model-state.json and inference.py, then run:

python inference.py --state model-state.json --post-id 2107943767382847844 --age-seconds 25200

This returns the fitted curve value. For a live cumulative forecast, anchor a nonnegative model increment at the latest observed count; do not describe the curve value as a verified future total.

Reproduce training and rolling-origin evaluation:

OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 python train.py --metrics training-aggregates.json --model torch --outputpath backtest-reproduced.json --model-outputpath state-reproduced.json

Python standard library handles loading and inference; PyTorch is needed for training. The state uses ordinary JSON rather than a pickle checkpoint. No inference endpoint or paid hosted job is deployed.

下一步

先收集同年龄、同口径的前瞻样本，保留简单基线；再逐步做「统计 → 加文本 → 加媒体 → 加评论」消融。只有在未参与选择的后续帖子上降低误差，才扩大模型。跨平台迁移复用数据合同、指标与回测流程，各平台单独校准。

训练材料仅含经授权的账号聚合指标和公共帖子链接。没有发布粉丝身份、个人质量判断、浏览器记录、凭据或私人素材。