推荐系统：下一本书怎么选Recommender System: Choose the Next Book
小禾借过两本科幻，借书机就推荐第三本。它知道小禾一定喜欢吗？After He borrows two science-fiction books, a library system suggests a third. Does it know he will like it?
第 1 步：先缩小候选范围Step 1: Find a manageable candidate set
书架太大，不能每本都仔细比较。There are too many books to compare equally.
推荐系统先从大量物品里找候选。可以看内容相似，也可以看用户互动模式。借过科幻，只说明一次行为。可能是替同学借，也可能并不喜欢。点击、停留和评分是线索，不是人的全部兴趣。新书和新用户还可能缺乏可用记录。A recommender finds candidates from a large collection using content similarity, interaction patterns or other signals. Borrowing science fiction is one behavior: he may be borrowing for a friend or may dislike the book. Clicks, time and ratings are clues rather than complete preferences. New items and users may have little history.
第 2 步：打分，再排序Step 2: Score, then rank
找到候选后，谁排第一？Which candidate goes first?
系统可把多种信号合成分数。比如主题相关、口碑和新鲜度。权重改变，排序也会改变。预测点击最高，不等于最值得学习。下方用公开的小公式展示排序。真实系统往往更复杂，还会处理重复、资格和多样性约束。A system can combine topic relevance, quality and freshness into scores. Changing weights can change the order. Highest predicted click rate is not necessarily greatest learning value. The experiment uses a small visible formula. Real systems often add complexity, eligibility rules, duplicate removal and diversity constraints.
第 3 步：看反馈，也留探索Step 3: Use feedback and exploration
只推熟悉主题，会漏掉什么？What is lost by recommending only familiar topics?
只看已经喜欢的内容，可能错过新兴趣。系统可留一些位置给探索。推荐结果又会影响下一轮行为，形成反馈环。小禾可以主动搜索、调整偏好。评估也应看多样性和长期满意度。没有一张推荐榜能代表所有人的最佳选择。Only familiar content may hide new interests. A system can reserve room for exploration. Recommendations influence later behavior, creating a feedback loop. He can search directly and adjust preferences. Evaluation can include diversity and longer-term satisfaction. No single ranking is best for everyone.
🎮 图书排序控制台🎮 Book ranking controls
调主题与新鲜度权重，让四份需求对应的书排到前面。公式和每本书的分数都可见。Adjust topic and freshness weights to satisfy four requests. The formula and book scores remain visible.
一句话记住它：Remember: 推荐是在特定目标下预测和排序，不是读心。Recommendations predict and rank for an objective; they do not read minds.
行为信号可能有多种解释。A behavior can have several explanations.改变权重会改变结果。Changing weights changes results.探索能带来新的信息。Exploration can reveal new information.
Back to tree
