csf123321/InsightRadar
1
0
Fork 0
mirror of https://github.com/stardrophere/InsightRadar.git synced 2026-06-05 23:07:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

算法与视觉优化

Browse Source
This commit is contained in:
stardrophere
2026-04-02 13:48:33 +08:00
parent 943770b2bc
commit 61b6357418
6 changed files with 113 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/app/services/fetcher_service.py
+2 -2
View File
@@ -26,9 +26,9 @@ SIMILARITY_THRESHOLD = float(os.getenv("SIMILARITY_THRESHOLD", 0.72))
API_BASE_URL = os.getenv("API_BASE_URL", "https://newsnow.busiyi.world/api/s")
EMBEDDING_MODEL_PATH = os.getenv("EMBEDDING_MODEL_PATH", "")
print("正在加载向量模型...")
print("正在加载 BAAI/bge-m3 向量模型...")
# 全局单例
embedder_model = SentenceTransformer(EMBEDDING_MODEL_PATH, local_files_only=True)
embedder_model = SentenceTransformer(EMBEDDING_MODEL_PATH, local_files_only=True, device="cuda")
print("模型加载完成。")
backend/app/services/matching_service.py
+82 -35
View File
@@ -1,6 +1,6 @@
"""
匹配服务：根据用户兴趣关键词（精确 + 语义）推荐事件
打分融合：匹配分 + 标签相关度 + 热度 + 新鲜度加成
打分融合：标签/标题匹配分 + 标签相关度 + 热度 + 新鲜度加成
"""
import os
from dataclasses import dataclass
@@ -14,7 +14,7 @@ from app.models.models import ExtractedTopic, TargetType, UnifiedEvent, UserTopi
from app.services.fetcher_service import embedder_model
# 语义匹配阈值：用户关键词和事件标签向量相似度达到该值才计入语义命中
# 语义匹配阈值：用户关键词和事件标签/标题向量相似度达到该值才计入语义命中
DEFAULT_PREFERENCE_SEMANTIC_THRESHOLD = 0.78
PREFERENCE_SEMANTIC_THRESHOLD = float(
os.getenv("PREFERENCE_SEMANTIC_THRESHOLD", str(DEFAULT_PREFERENCE_SEMANTIC_THRESHOLD))
@@ -41,6 +41,31 @@ def _normalize_text(text: str) -> str:
return text.strip().casefold()
def _find_exact_preference_match(
target_text: str,
normalized_preferences: list[tuple[str, str]],
) -> str | None:
"""
判断目标文本是否与某个用户兴趣词形成“精确命中”。
命中条件：
1. 标准化后完全相等
2. 二者互为包含关系
返回命中的原始兴趣词，未命中则返回 None。
"""
normalized_target = _normalize_text(target_text)
if not normalized_target:
return None
for raw_pref, normalized_pref in normalized_preferences:
if not normalized_pref:
continue
if normalized_target == normalized_pref:
return raw_pref
if normalized_pref in normalized_target or normalized_target in normalized_pref:
return raw_pref
return None
_EMBEDDING_CACHE: dict[str, np.ndarray] = {}
MAX_CACHE_SIZE = 10000
@@ -86,6 +111,26 @@ def _build_keyword_embedding_map(keywords: list[str]) -> dict[str, np.ndarray]:
return result
def _find_best_semantic_match(
target_text: str,
target_vec_map: dict[str, np.ndarray],
pref_vec_map: dict[str, np.ndarray],
) -> tuple[str | None, float]:
"""返回与目标文本最接近的兴趣词及其余弦相似度。"""
target_vec = target_vec_map.get(target_text)
if target_vec is None:
return None, -1.0
best_pref = None
best_sim = -1.0
for pref_keyword, pref_vec in pref_vec_map.items():
sim = float(np.dot(target_vec, pref_vec))
if sim > best_sim:
best_sim = sim
best_pref = pref_keyword
return best_pref, best_sim
def _ensure_aware(dt: datetime) -> datetime:
"""SQLite 读出的 datetime 不带时区信息，统一补上 UTC 后才能和 utcnow() 做减法。"""
if dt.tzinfo is None:
@@ -116,8 +161,8 @@ def recommend_events_for_user(
) -> list[MatchedEventResult]:
"""
用户兴趣推荐主流程：
1) 精确匹配：用户词 == EVENT 标签
2) 语义匹配：用户词向量 vs EVENT 标签向量（超过阈值）
1) 精确匹配：用户词 vs EVENT 标签/标题
2) 语义匹配：用户词向量 vs EVENT 标签/标题向量（超过阈值）
3) 打分融合：匹配分 + 标签相关度 + 热度 + 新鲜度
"""
final_limit = max(1, min(limit, PREFERENCE_RECOMMEND_MAX_LIMIT))
@@ -167,8 +212,6 @@ def recommend_events_for_user(
)
.all()
)
if not topic_rows:
return []
# 组织事件标签映射：event_id -> [(tag, relevance_score), ...]
event_topics: dict[int, list[tuple[str, float | None]]] = {}
@@ -177,10 +220,6 @@ def recommend_events_for_user(
continue
event_topics.setdefault(event_id, []).append((topic_keyword, relevance_score))
# 如果某事件没有标签，就不参与推荐
if not event_topics:
return []
# 3. 批量编码用户词与标签词，减少模型调用次数
unique_preference_keywords = list(dict.fromkeys(preference_keywords))
unique_topic_keywords = list(dict.fromkeys([row[1] for row in topic_rows if row[1]]))
@@ -188,13 +227,21 @@ def recommend_events_for_user(
topic_vec_map = _build_keyword_embedding_map(unique_topic_keywords)
# 预先建立“标准化后用户词集合”，用于精确匹配
normalized_pref_set = {_normalize_text(word) for word in unique_preference_keywords}
normalized_preference_pairs = [
(word, _normalize_text(word))
for word in unique_preference_keywords
if _normalize_text(word)
]
unique_event_titles = list(
dict.fromkeys(
[event.unified_title.strip() for event in events if event.unified_title and event.unified_title.strip()]
)
)
title_vec_map = _build_keyword_embedding_map(unique_event_titles)
scored_results: list[MatchedEventResult] = []
for event in events:
topic_list = event_topics.get(event.id, [])
if not topic_list:
continue
exact_hits: list[str] = []
semantic_hits: list[dict[str, Any]] = []
@@ -202,37 +249,18 @@ def recommend_events_for_user(
# 对每个事件标签做精确匹配或语义匹配
for topic_keyword, topic_relevance in topic_list:
normalized_topic = _normalize_text(topic_keyword)
topic_relevance_score = float(topic_relevance) if topic_relevance is not None else 50.0
# 1) 精确命中（包括完全相等与包含关系）
matched_exact = False
if normalized_topic in normalized_pref_set:
matched_exact = True
else:
for pref_word in normalized_pref_set:
if pref_word and (pref_word in normalized_topic or normalized_topic in pref_word):
matched_exact = True
break
if matched_exact:
matched_pref = _find_exact_preference_match(topic_keyword, normalized_preference_pairs)
if matched_pref is not None:
exact_hits.append(topic_keyword)
# 精确命中给较高基础分，标签自身相关度作为增益
score += 45.0 + topic_relevance_score * 0.2
continue
# 2) 语义命中（未精确命中时再算）
topic_vec = topic_vec_map.get(topic_keyword)
if topic_vec is None:
continue
best_pref = None
best_sim = -1.0
for pref_keyword, pref_vec in pref_vec_map.items():
sim = float(np.dot(topic_vec, pref_vec))
if sim > best_sim:
best_sim = sim
best_pref = pref_keyword
best_pref, best_sim = _find_best_semantic_match(topic_keyword, topic_vec_map, pref_vec_map)
if best_pref is not None and best_sim >= similarity_threshold:
semantic_hits.append(
@@ -245,6 +273,25 @@ def recommend_events_for_user(
# 语义命中分略低于精确命中，并由相似度放大
score += best_sim * 35.0 + topic_relevance_score * 0.12
# 标题也参与匹配，但权重低于结构化标签，避免长标题过度主导排序。
event_title = (event.unified_title or "").strip()
if event_title:
title_exact_pref = _find_exact_preference_match(event_title, normalized_preference_pairs)
if title_exact_pref is not None:
exact_hits.append(f"标题:{title_exact_pref}")
score += 30.0
else:
best_pref, best_sim = _find_best_semantic_match(event_title, title_vec_map, pref_vec_map)
if best_pref is not None and best_sim >= similarity_threshold:
semantic_hits.append(
{
"preference_keyword": best_pref,
"topic_keyword": f"标题:{best_pref}",
"similarity": round(best_sim, 4),
}
)
score += best_sim * 24.0
# 如果精确和语义都没命中，直接跳过
if not exact_hits and not semantic_hits:
continue
frontend/src/components/UnifiedEventCard.vue
+8
View File
@@ -111,6 +111,14 @@ function getRankingChartOptions(history: number[], platformColor: string) {
height: 56,
sparkline: { enabled: true },
animations: { enabled: true, easing: 'easeinout' as const, speed: 400 },
events: {
mounted: (chartContext: any) => {
chartContext.el?.querySelector('.apexcharts-svg > title')?.remove()
},
updated: (chartContext: any) => {
chartContext.el?.querySelector('.apexcharts-svg > title')?.remove()
}
}
},
stroke: { curve: 'smooth' as const, width: 2 },
fill: {
frontend/src/views/DashboardView.vue
+8
View File
@@ -182,6 +182,14 @@ function getRankingChartOptions(history: number[], platformColor: string) {
height: 56,
sparkline: { enabled: true },
animations: { enabled: true, easing: 'easeinout' as const, speed: 400 },
events: {
mounted: (chartContext: any) => {
chartContext.el?.querySelector('.apexcharts-svg > title')?.remove()
},
updated: (chartContext: any) => {
chartContext.el?.querySelector('.apexcharts-svg > title')?.remove()
}
}
},
stroke: { curve: 'smooth' as const, width: 2 },
fill: {
frontend/src/views/SearchView.vue
+12 -2
View File
@@ -72,6 +72,12 @@ const chartOptions = ref<ApexOptions>({
},
// 点击图表数据点：切换选中时间，再次点击则取消筛选
events: {
mounted: (chartContext: any) => {
chartContext.el?.querySelector('.apexcharts-svg > title')?.remove()
},
updated: (chartContext: any) => {
chartContext.el?.querySelector('.apexcharts-svg > title')?.remove()
},
markerClick: function(event: unknown, chartContext: unknown, { dataPointIndex }: never) {
if (searchResult.value && searchResult.value.timeline[dataPointIndex]) {
const clickedTime = searchResult.value.timeline[dataPointIndex].time_label
@@ -585,7 +591,12 @@ async function handleSearch() {
.chart-container {
margin-top: 16px;
margin-left: -10px; /* 视觉上抵消 apexcharts 的默认左侧留白。 */
margin-left: -10px;
}
.chart-container :deep(svg),
.chart-container :deep(canvas) {
outline: none;
}
.events-section {
@@ -595,7 +606,6 @@ async function handleSearch() {
.events-grid {
display: flex;
flex-direction: column;
/* 与 DashboardView 保持一致，列表按纵向堆叠展示。 */
}
.loading-state {
frontend/src/views/TopicsView.vue
+1 -1
View File
@@ -156,7 +156,7 @@ onMounted(async () => {
v-model="newKeyword"
type="text"
class="keyword-input"
placeholder="输入关键词，如「直升机」「科比」「佐巴扬」..."
placeholder="输入关键词，如「篮球」「科比」「科技」..."
maxlength="100"
@keydown="onInputKeydown"
/>