| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320 |
- """
- 条件概率计算工具:
- 1)计算某个人设树节点在父节点下的条件概率;
- 2)计算某个 pattern 的条件概率。
- """
- from __future__ import annotations
- import itertools
- import json
- from pathlib import Path
- from typing import Any
- # 节点名 -> (该节点 post_ids, 父节点 post_ids),用 frozenset 便于批量计算时复用、避免重复转换
- NodePostIndex = dict[str, tuple[frozenset[str], frozenset[str]]]
- # 已推导列表:每项为 (已推导的选题点, 推导来源人设树节点),如 ("分享","分享")、("柴犬","动物角色")
- # 推导来源人设树节点的 post_ids 在计算条件概率时从人设树中读取
- DerivedItem = tuple[str, str]
- def _tree_dir(account_name: str, base_dir: Path | None = None) -> Path:
- """人设树目录:../input/{account_name}/处理后数据/tree/(相对本文件所在目录)。"""
- if base_dir is not None:
- return base_dir / account_name / "处理后数据" / "tree"
- return Path(__file__).resolve().parent.parent / "input" / account_name / "处理后数据" / "tree"
- def _load_trees(account_name: str, base_dir: Path | None = None) -> list[tuple[str, dict]]:
- """加载该账号下所有维度的人设树。返回 [(维度名, 根节点 dict), ...]。"""
- td = _tree_dir(account_name, base_dir)
- if not td.is_dir():
- return []
- return _load_trees_from_directory(td)
- def _load_trees_from_directory(tree_dir: Path) -> list[tuple[str, dict]]:
- """
- 从指定目录加载所有人设树 JSON(每文件取顶层第一个维度根,与按账号目录加载时行为一致)。
- 用于平台库等人设树路径非 input/{账号}/处理后数据/tree/ 的场景。
- """
- if not tree_dir.is_dir():
- return []
- result: list[tuple[str, dict]] = []
- for p in sorted(tree_dir.glob("*.json")):
- try:
- with open(p, "r", encoding="utf-8") as f:
- data = json.load(f)
- for dim_name, root in data.items():
- if isinstance(root, dict):
- result.append((str(dim_name), root))
- break
- except Exception:
- continue
- return result
- def _post_ids_of(node: dict) -> list[str]:
- """从树节点中取出 _post_ids,无则返回空列表。"""
- return list(node.get("_post_ids") or [])
- def _build_node_index_from_trees(trees: list[tuple[str, dict]]) -> dict[str, tuple[list[str], list[str]]]:
- """
- 遍历多棵人设树,建立 节点名 -> (该节点 post_ids, 父节点 post_ids)。
- 同一节点名在多个分支出现时,保留第一次遇到的(保证父子一致)。
- """
- index: dict[str, tuple[list[str], list[str]]] = {}
- for _dim, root in trees:
- parent_pids = _post_ids_of(root)
- def walk(parent_ids: list[str], node_dict: dict) -> None:
- for name, child in (node_dict.get("children") or {}).items():
- if not isinstance(child, dict):
- continue
- if name not in index:
- index[name] = (_post_ids_of(child), list(parent_ids))
- walk(_post_ids_of(child), child)
- walk(parent_pids, root)
- return index
- def _build_node_index(account_name: str, base_dir: Path | None = None) -> dict[str, tuple[list[str], list[str]]]:
- """遍历账号下所有人设树,建立节点索引。"""
- return _build_node_index_from_trees(_load_trees(account_name, base_dir))
- def build_node_post_index_from_tree_dir(tree_dir: Path) -> NodePostIndex:
- """从任意人设树目录(如 input/xiaohongshu/tree)构建节点 post 索引,算法与账号树一致。"""
- raw = _build_node_index_from_trees(_load_trees_from_directory(tree_dir))
- return {k: (frozenset(a), frozenset(b)) for k, (a, b) in raw.items()}
- def build_node_index_for_tree_dir(tree_dir: Path) -> dict[str, tuple[list[str], list[str]]]:
- """从任意人设树目录构建节点名 -> (节点 post_ids, 父 post_ids),供 pattern 条件概率等使用。"""
- return _build_node_index_from_trees(_load_trees_from_directory(tree_dir))
- def load_persona_trees_from_dir(tree_dir: Path) -> list[tuple[str, dict]]:
- """从目录加载人设树列表(每 JSON 文件一个顶层维度),供遍历节点等场景复用。"""
- return _load_trees_from_directory(tree_dir)
- def _derived_post_ids_from_sources(
- derived_list: list[DerivedItem],
- index: dict[str, tuple[list[str], list[str]]],
- ) -> set[str]:
- """根据 derived_list 中的「推导来源人设树节点」在人设树中的 post_ids 取交集,得到已推导的帖子集合。"""
- common: set[str] | None = None
- for _topic_point, source_node in derived_list:
- if source_node not in index:
- continue
- pids = set(index[source_node][0])
- if common is None:
- common = pids
- else:
- common &= pids
- return common if common is not None else set()
- def _derived_post_ids_from_frozen_index(
- derived_list: list[DerivedItem],
- index: NodePostIndex,
- ) -> frozenset[str]:
- """与 _derived_post_ids_from_sources 相同语义,索引为 frozenset 版(批量场景复用)。"""
- common: frozenset[str] | None = None
- for _topic_point, source_node in derived_list:
- if source_node not in index:
- continue
- pids = index[source_node][0]
- common = pids if common is None else common & pids
- return common if common is not None else frozenset()
- def build_node_post_index(account_name: str, base_dir: Path | None = None) -> NodePostIndex:
- """
- 构建账号人设树的节点索引(每个节点只建一次,供批量 calc_node_conditional_ratio 复用)。
- 值为 (节点 post_ids, 父节点 post_ids) 的 frozenset,减少重复 list->set 与拷贝。
- """
- raw = _build_node_index(account_name, base_dir)
- return {k: (frozenset(a), frozenset(b)) for k, (a, b) in raw.items()}
- def calc_node_conditional_ratio(
- account_name: str,
- derived_list: list[DerivedItem],
- tree_node_name: str,
- base_dir: Path | None = None,
- node_post_index: NodePostIndex | None = None,
- target_ratio: float | None = None,
- ) -> float:
- """
- 计算人设树节点 N 在父节点 P 下的条件概率。
- 参数:
- account_name: 账号名称
- derived_list: 已推导列表,每项 (已推导的选题点, 推导来源人设树节点)
- tree_node_name: 人设树节点 N 的名称(字符串匹配)
- base_dir: 可选,input 根目录;不传则使用相对本文件的 ../input
- node_post_index: 可选,由 build_node_post_index 预构建;批量对多节点计算时传入可避免重复读盘与遍历整棵树
- target_ratio: 可选,目标条件概率。若某个组合的条件概率已达到该值,则直接返回(用于缩小组合搜索)
- 计算规则:
- 已推导的帖子集合:从 derived_list 中先取「最多选题点」的交集,再逐步减少到 1 个选题点,
- 对每种选题点子集分别计算条件概率,最后取最大值。
- 对每种情况:已推导的帖子集合 = 该子集中各「推导来源人设树节点」在人设树中的 post_ids 的交集;
- 分子 = |已推导的帖子集合 ∩ N 的 post_ids|,分母 = |已推导的帖子集合 ∩ P 的 post_ids|;
- 条件概率 = 分子/分母,且 ≤1;分母为 0 时该情况跳过。
- """
- index = node_post_index if node_post_index is not None else build_node_post_index(account_name, base_dir)
- if tree_node_name not in index:
- return 0.0
- set_n, set_p = index[tree_node_name]
- # 关键优化(不改变搜索空间/结果):
- # - derived_list 里重复的 source_node 对“交集”没有任何影响,但会把 L 变大导致 2^L 爆炸
- # - 不在 index 里的 source_node 原本也会被跳过,提前过滤可减少组合规模
- # - 组合内交集直接对 frozenset 逐步 &,避免 list(combo)/函数调用开销
- seen_sources: set[str] = set()
- source_sets: list[frozenset[str]] = []
- for _topic, source_node in derived_list:
- if source_node in seen_sources:
- continue
- seen_sources.add(source_node)
- tup = index.get(source_node)
- if tup is None:
- continue
- source_sets.append(tup[0])
- if not source_sets:
- return 0.0
- # 将更小的集合放前面:交集会更快“变小”,每次 & 的成本更低(仍然枚举全部子集)
- source_sets.sort(key=len)
- max_ratio = 0.0
- # 从 1 个选题点到「最多选题点」:对每种子集大小,取所有组合,分别算条件概率后取最大
- for k in range(1, len(source_sets) + 1):
- for combo_sets in itertools.combinations(source_sets, k):
- derived_post_ids = combo_sets[0]
- for s in combo_sets[1:]:
- derived_post_ids = derived_post_ids & s
- den = len(derived_post_ids & set_p)
- if den == 0:
- continue
- num = len(derived_post_ids & set_n)
- ratio = min(1.0, num / den)
- max_ratio = max(max_ratio, ratio)
- if target_ratio is not None and max_ratio >= target_ratio:
- return round(max_ratio, 4)
- return round(max_ratio, 4)
- def _pattern_nodes_and_post_count(pattern: dict[str, Any]) -> tuple[list[str], int, float]:
- """
- 从 pattern 中解析出节点列表和 post_count。支持 nodes + post_count 或 i + post_count。
- 返回的 post_count 表示该 pattern 本身的帖子数,在条件概率计算中作为分子(即 pattern 本身的概率/占比的分子)。
- """
- nodes = pattern.get("nodes")
- if nodes is not None and isinstance(nodes, list):
- nodes = [str(x).strip() for x in nodes if x]
- else:
- raw = pattern.get("i") or pattern.get("pattern_str") or ""
- nodes = [x.strip() for x in str(raw).replace("+", " ").split() if x.strip()]
- post_count = int(pattern.get("post_count", 0))
- support = pattern.get("s", 0.0)
- return nodes, post_count, support
- def calc_pattern_conditional_ratio_with_index(
- derived_list: list[DerivedItem],
- pattern: dict[str, Any],
- index: dict[str, tuple[list[str], list[str]]],
- ) -> float:
- """
- 与 calc_pattern_conditional_ratio 相同计算规则,但使用已构建的人设树节点索引
- (例如平台库 input/xiaohongshu/tree)。
- """
- pattern_nodes, post_count, pattern_s = _pattern_nodes_and_post_count(pattern)
- if not pattern_nodes or post_count <= 0:
- return pattern_s
- derived_sources = set(source for _post, source in derived_list)
- derived_pattern_nodes = [n for n in pattern_nodes if n in derived_sources]
- if not derived_pattern_nodes:
- return pattern_s
- derived_in_tree = [n for n in derived_pattern_nodes if n in index]
- if not derived_in_tree:
- return pattern_s
- common: set[str] | None = None
- for name in derived_in_tree:
- pids = set(index[name][0])
- if common is None:
- common = pids
- else:
- common &= pids
- if common is None or len(common) == 0:
- return pattern_s
- den = len(common)
- return round(min(1.0, post_count / den), 4)
- def calc_pattern_conditional_ratio(
- account_name: str,
- derived_list: list[DerivedItem],
- pattern: dict[str, Any],
- base_dir: Path | None = None,
- ) -> float:
- """
- 计算某个 pattern 的条件概率。
- 参数:
- account_name: 账号名称
- derived_list: 已推导列表,每项 (已推导的选题点, 推导来源人设树节点)
- pattern: 至少包含节点列表与 post_count。
- - 节点列表: key 为 "nodes"(list)或 "i"(字符串,用 + 连接)
- - post_count: 该 pattern 的帖子数量,作为分子
- base_dir: 可选,input 根目录
- 计算规则:
- 取 pattern 中「已被推导」的节点(其名称出现在 derived 的推导来源中),
- 在人设树中取这些节点的 post_ids 的交集作为分母;
- 分子 = pattern.post_count(由 _pattern_nodes_and_post_count 解析得到,表示 pattern 本身的帖子数)。
- 条件概率 = 分子/分母,且 ≤1;分母为 0 时返回 1。
- """
- index = _build_node_index(account_name, base_dir)
- return calc_pattern_conditional_ratio_with_index(derived_list, pattern, index)
- def _test_with_user_example() -> None:
- """
- 使用你提供的测试数据:已推导 (分享|分享)、(柴犬|动物角色);
- 人设树节点:恶作剧;pattern:分享+动物角色+创意表达 post_count=2。
- 推导来源的 post_ids 在方法内部从人设树读取。
- """
- account_name = "阿里多多酱"
- # 已推导列表:(已推导的选题点, 推导来源人设树节点)
- derived_list: list[DerivedItem] = [
- ("推广", "推广"),
- ("视觉调性", "视觉调性"),
- # ("图片文字", "图片文字"),
- # ("补充说明式", "补充说明式"),
- # ("幽默化标题", "幽默化标题"),
- # ("标题", "标题"),
- ]
- # 1)人设树节点「恶作剧」的条件概率
- r_node = calc_node_conditional_ratio(account_name, derived_list, "观念")
- print(f"1) 人设树节点条件概率: {r_node}")
- # 2)pattern 分享+动物角色+创意表达 post_count=2 的条件概率
- pattern = {"i": "视觉调性+辞格意象+叙事编排", "post_count": 22, "s": 0.478261}
- r_pattern = calc_pattern_conditional_ratio(account_name, derived_list, pattern)
- print(f"2) pattern 条件概率: {r_pattern}")
- if __name__ == "__main__":
- _test_with_user_example()
|