# StreamProfileBench

Streaming user-interest profiling benchmark for Chinese social media platforms.
Each task asks an LLM to (a) maintain a rolling persona summary from a stream
of a user's posts, and (b) predict which tags from a curated candidate pool the
user will engage with in the next time window.

The benchmark evaluates two coupled abilities:

- **Plasticity** — picking up *newly emerging* interests (`Recall_Novelty`).
- **Stability** — retaining *persisting* interests (`Recall_Stability`).

It also probes robustness against four classes of distractors (peer-cluster,
viral, decayed, random), and forward transfer from accumulated personas.

## Repository layout

```
streamprofile_bench_release/
├── bench_inference.py    # main runner: inference + scoring (and --reeval)
├── bench_task.py         # prompt formatter (PLATFORM_CONTEXT + format_prompt)
├── stats.py              # aggregate eval reports across models into Excel
├── sample_subset.py      # sample N% of users per platform for cheap dry-runs
├── data/
│   ├── weibo_inference_tasks.jsonl
│   ├── xiaohongshu_inference_tasks.jsonl
│   ├── toutiao_inference_tasks.jsonl
│   ├── zhihu_inference_tasks.jsonl
│   └── douban_inference_tasks.jsonl
├── requirements.txt
├── LICENSE                # Apache-2.0
└── README.md
```

## Install

```bash
pip install -r requirements.txt
```

Python 3.9+.

## Quick start

```bash
export LLM_API_KEY="sk-..."
export LLM_API_BASE="https://api.openai.com/v1"   # any OpenAI-compatible endpoint

# Single platform
python bench_inference.py --model gpt-4o-mini --platform weibo

# All five platforms
python bench_inference.py --model gpt-4o-mini
```

Outputs go to `results/<model>/`:
- `inference_results_<platform>.jsonl` — per-user, per-step predictions, ground truth, persona summaries, and metrics.
- `eval_report_<platform>.txt` — text report (`M_bar`, `F1^NS`, `FWT`, cold-start vs persona-augmented, learning curve).

### Re-evaluate without re-running the model

If you change metric definitions or want to recompute scores from saved
predictions, no API calls needed:

```bash
python bench_inference.py --model gpt-4o-mini --reeval
python bench_inference.py --model gpt-4o-mini --platform weibo --reeval
```

This rewrites the `metrics` field in each `inference_results_<platform>.jsonl`
and produces a fresh `eval_report_<platform>.txt`.

### Cheap dry-run on a 20% subset

```bash
python sample_subset.py --ratio 0.2
# produces data/<platform>_inference_tasks_sub20.jsonl
```

To use the subset, point `bench_inference.py` at the sub-files (e.g. by
renaming, symlinking, or temporarily editing `DATA_DIR`).

### Aggregate results across models into Excel tables

```bash
python stats.py
# writes metric_excels/{Precision,Recall,F1^NS,...}.xlsx
```

Each Excel file has one row per model, one column per platform (plus an
`Avg.` column).

## Configuration

| Variable               | Purpose                                                       | Default |
|------------------------|---------------------------------------------------------------|---------|
| `LLM_API_KEY`          | API key for your LLM provider                                 | (required) |
| `LLM_API_BASE`         | OpenAI-compatible base URL                                    | (required) |
| `LLM_INSECURE_TLS`     | Set `1` to skip TLS verification (e.g. for self-signed dev)   | `0` |
| `BENCH_MAX_WORKERS`    | Parallel inference threads                                    | `16` |
| `--model`              | Model name passed to the API                                  | (required) |
| `--api_url`/`--api_key`| Override the env vars                                         | unset |
| `--platform`           | One of `weibo / xiaohongshu / toutiao / zhihu / douban`       | all |
| `--reeval`             | Skip inference; re-score saved predictions                    | off |

## Data format

One JSON object per line in each `data/<platform>_inference_tasks.jsonl`:

```json
{
  "user_id": "we_16d82b9a92",
  "platform": "weibo",
  "username": "U_010ada10",
  "bio": "...",
  "total_steps": 4,
  "prediction_tasks": [
    {
      "step_id": 1,
      "total_steps": 4,
      "date_input":  "2025-06-12",
      "date_target": "2025-06-17",
      "posts_text":  "[1] Content: ...\nTags: ...\n[2] ...",
      "candidate_pool": ["tag1", "tag2", ...],
      "ground_truth": {
        "all_tags":  ["..."],
        "new_tags":  ["..."],
        "keep_tags": ["..."]
      },
      "meta": {
        "T_keep":   ["..."],
        "T_new":    ["..."],
        "D_decay":  ["..."],
        "D_cluster":["..."],
        "D_viral":  ["..."],
        "D_random": ["..."]
      }
    }
  ]
}
```

Per platform / step the candidate pool is pre-built as
`pool_size = clip(|GT|·4, [10, 50])` with the remaining slots filled with
distractors of four types:

- **D_decay**   — tags the user engaged with in the *current* batch but does **not** carry forward (interests on the way out).
- **D_cluster** — tags from the same semantic cluster as the GT (peer-cluster distractors).
- **D_viral**   — tags trending platform-wide on the target date (popular but irrelevant).
- **D_random**  — tags sampled uniformly from the platform's tag vocabulary.

Ground truth is `T_new ∪ T_keep`, where `T_keep = current ∩ future` and
`T_new = future \ current`.

## Metrics

| Metric              | Definition                                                                 |
|---------------------|----------------------------------------------------------------------------|
| **Precision**       | `|pred ∩ GT| / |pred|`                                                     |
| **Recall**          | `|pred ∩ GT| / |GT|`                                                       |
| **Recall_Novelty**  | Recall restricted to `T_new` (plasticity)                                  |
| **Recall_Stability**| Recall restricted to `T_keep` (stability)                                  |
| **F1^NS**           | Harmonic mean of `Recall_Novelty` and `Recall_Stability`, per user, then macro-averaged |
| **Error_Peer**      | False-positive rate on `D_cluster` (lower is better)                       |
| **Error_Viral**     | False-positive rate on `D_viral`                                           |
| **Error_Decay**     | False-positive rate on `D_decay`                                           |
| **Error_Random**    | False-positive rate on `D_random`                                          |
| **FWT**             | Forward transfer: `mean(M_{t≥2}) − M_{t=1}`, per metric                    |

Aggregation in the report:

- **`M_bar`** — within-user mean across steps, then macro mean across users.
- **Cold-start vs persona-augmented** — same aggregation on `step_id == 1`
  vs `step_id ≥ 2`, plus the delta. Quantifies the value of carrying personas.
- **Learning curve** — per `step_id` position, mean across users.

## Anonymization notes

This release uses anonymized identifiers:

- `user_id` is a hash prefixed with the platform code (`we_` weibo, `xi_` xiaohongshu, `to_` toutiao, `zh_` zhihu, `do_` douban).
- `username` values are replaced with placeholders of the form `U_xxxxxxx`.
- Free-text fields (`bio`, `posts_text`) retain user-generated content as
  collected; we do not redact post bodies.

If you find any residual PII please open an issue.

## Citation

```bibtex
@misc{streamprofilebench2026,
  title  = {StreamProfileBench: Streaming User-Interest Profiling on Chinese Social Media},
  author = {...},
  year   = {2026},
  howpublished = {GitHub},
  url    = {...}
}
```

## License

Apache License 2.0. See [LICENSE](LICENSE).