YouTube Randomizer

Ranking integrity update with exposure snapshots, attempt IDs, shared swap gates, canonical supply, and final validation.

開發者介紹 / Developer Introduction 最近發現 YouTube 首頁推薦的內容越來越容易重複，相似影片、相同頻道和相近主題會反覆出現，因此開發了 YouTube Randomizer。它會在 YouTube 原本的推薦結果上加入一層「探索機制」，讓首頁不只停留在原本的推薦範圍，而是隨著你的操作逐步向外探索新的影片、頻道與主題。 一句話介紹 如果這一頁沒有想看的內容，就點 YouTube Logo 或按 F5。每刷新一次，Randomizer 都會逐步擴大探索範圍，直到你找到一支真正願意看一段時間的影片。 同時，Randomizer 會記錄你實際看過的影片、觀看時間、近期推薦內容與重複曝光情況，逐步調整後續排序，讓推薦方向更接近你的實際使用習慣。 所有資料都只保存在瀏覽器本機，不需要額外帳號，也不會傳送到外部伺服器。 I noticed that YouTube's homepage recommendations can become increasingly repetitive, with similar videos, the same channels, and closely related topics appearing again and again. That's why I built YouTube Randomizer. It adds an additional discovery layer on top of YouTube's existing recommendations, allowing the homepage to gradually explore beyond its usual recommendation range and surface different videos, channels, and topics. In One Sentence If you don't see anything you want to watch, click the YouTube logo or press F5. Each refresh gradually expands the discovery range until you find a video you actually want to spend some time watching. Randomizer also learns from videos you actually watch, watch duration, recent recommendations, and repeated exposure, then gradually adjusts future rankings based on your usage behavior. All data is stored locally in your browser. No additional account or external server is required. 演算法運作方式 / How the Algorithm Works YouTube Randomizer 使用本地端 Discovery Algorithm，在 YouTube 原本推薦內容的基礎上重新調整首頁排序。它會參考： 近期首頁曾出現過哪些影片 同一影片與頻道的重複曝光次數 實際點擊與觀看行為 觀看時間與觀看比例 近期興趣變化 YouTube 原本的推薦順序 本輪刷新期間已經出現過的影片、頻道與主題 當你持續重新整理首頁卻沒有找到想看的影片時，Randomizer 會逐步增加探索強度。前期主要優先更換影片，中期開始增加不同頻道與相鄰主題，後期則會更積極探索新的內容方向。 探索範圍會逐步從： 不同影片 → 不同頻道 → 相鄰主題 → 更廣泛的新主題 向外擴張。 Randomizer 不會完全捨棄 YouTube 原本的推薦結果，而是利用 YouTube 已經提供的候選影片進行重新排序，因此即使進入較高探索階段，仍會保留一定程度的相關性，避免首頁變成完全隨機內容。 如果某支影片或某個頻道在同一輪探索中反覆出現，它的排序權重會逐漸下降；反之，本輪尚未出現過的新影片、新頻道與新主題則會逐步提高曝光機會。 當系統偵測到你確實點進某支影片並觀看一段時間後，代表這一輪探索已經找到有效內容。下一次回到首頁時，探索範圍會重新收斂，從較接近正常 YouTube 推薦的狀態重新開始。 YouTube Randomizer uses a local Discovery Algorithm to re-rank videos already provided by YouTube. It considers factors such as: Recently displayed recommendations Repeated exposure of the same videos and channels Actual clicks and watch behavior Watch duration and watch ratio Recent changes in interests YouTube's original recommendation order Videos, channels, and topics already shown during the current refresh session If you repeatedly refresh the homepage without finding something you want to watch, Randomizer gradually increases its exploration range. Early refreshes mainly rotate videos. As the refresh streak increases, the algorithm starts introducing more channels, adjacent topics, and eventually broader discovery directions. The exploration path gradually expands from: Different videos → Different channels → Adjacent topics → Broader new topics Randomizer does not completely discard YouTube's original recommendations. It still uses YouTube's candidate pool as the foundation, so even during deeper exploration, content retains a degree of relevance instead of becoming completely random. Videos or channels that repeatedly appear during the same exploration session are gradually down-ranked, while unseen videos, new channels, and new topic directions receive more exposure. Once you open a video and watch it long enough for the system to recognize it as meaningful viewing, the current exploration session is considered successful. The next time you return to the homepage, the exploration range contracts and starts again from a state closer to YouTube's normal recommendations. 探索機制 / Discovery Stages 目前的探索流程大致分成幾個階段。初始階段會盡量保留 YouTube 原本的推薦方向，只進行小幅度重排。連續刷新後，會先提高「新影片」的比例，避免一直看到相同內容。如果仍然沒有找到想看的影片，系統會開始增加不同頻道的曝光，降低相同頻道重複出現的頻率。再繼續刷新後，演算法會開始往相鄰主題擴張，例如從一個興趣的子類別，延伸到相關或可能有交集的內容。在更高的探索階段，Randomizer 會逐步增加較遠的新主題與新方向，但仍保留相關性判斷與安全限制，不會直接變成完全隨機推薦。 簡化來看： 刷新越多次 → 本輪重複內容越少 → 新影片越多 → 新頻道越多 → 主題距離逐步擴大 The discovery process is divided into several levels. At the beginning, Randomizer stays close to YouTube's original recommendation direction and only makes small ranking adjustments. After repeated refreshes, it prioritizes unseen videos to reduce repetition. If you still do not find anything worth watching, the algorithm begins increasing channel diversity and reducing repeated exposure from the same creators. With additional refreshes, exploration expands into adjacent topics and related areas. At higher exploration levels, Randomizer becomes more willing to test broader and less familiar topics, while still keeping relevance safeguards in place. In short: More refreshes → Less repetition → More new videos → More new channels → Wider topic exploration 其他功能 / Additional Features 降低重複影片與頻道的出現頻率 增加新影片與新頻道曝光 根據實際觀看行為調整興趣方向 支援 Shorts、不同影片來源、倍速播放與背景播放判斷 支援多興趣與多語言使用情境 保留部分隨機性，降低形成新推薦同溫層的機率 提供 ON / OFF 開關 關閉後恢復 YouTube 原本排序 所有學習紀錄只保存在瀏覽器本機 不使用定位 不需要額外帳號 不需要外部伺服器 Reduces repeated videos and channels Increases exposure to new videos and creators Adjusts interests based on actual watch behavior Supports Shorts, different video sources, playback speed, and background playback detection Supports users with multiple interests and languages Retains limited randomness to reduce the risk of creating a new recommendation bubble ON / OFF toggle Restores YouTube's original ordering when disabled All learning data is stored locally No location tracking No additional account required No external server required 使用方式 / How to Use 安裝擴充功能後，開啟 YouTube 首頁。 YouTube Logo 旁會出現 ON / OFF 開關。 設為 ON 後，如果首頁沒有想看的影片，可以點擊 YouTube Logo 或按 F5，重新取得一批推薦內容。 只要這一輪還沒有找到有效觀看的影片，每次重新取得首頁都會逐步擴大探索範圍。 當你找到影片並實際觀看一段時間後，系統會確認這一輪探索成功。下一次回到首頁時，探索強度會重新回到較低階段。 設為 OFF 時，Randomizer 不會修改 YouTube 原本的推薦排序。 ON / OFF 狀態會自動保存在瀏覽器中。 After installing the extension, open the YouTube homepage. An ON / OFF toggle will appear next to the YouTube logo. When set to ON, if you don't see anything you want to watch, click the YouTube logo or press F5 to get a new set of recommendations. As long as the current exploration session has not found meaningful viewing, each refresh gradually expands the discovery range. Once you find a video and actually spend some time watching it, Randomizer recognizes the exploration session as successful. The next time you return to the homepage, exploration starts again from a lower level. When set to OFF, Randomizer leaves YouTube's original recommendation order unchanged. The ON / OFF state is saved automatically in your browser. 注意 / Note：切換 ON / OFF 不會立即重新排列目前已經顯示的內容，新狀態會在下一次重新整理、點擊 YouTube Logo、返回首頁或重新進入 YouTube 首頁時生效。關閉 Randomizer 不會清除已建立的長期興趣紀錄，但會結束目前這一輪 Refresh Session。 Switching ON / OFF does not immediately rearrange content already displayed on the current page. The new state takes effect the next time you refresh, click the YouTube logo, return to the homepage, or re-enter YouTube. Turning Randomizer OFF does not erase the long-term interest profile, but it ends the current Refresh Session. 更新日誌 / Changelog 第一次更新 / First Update 中文 完成 YouTube Randomizer 基礎版本 加入本地端首頁推薦重新排序功能，開始記錄近期推薦影片與頻道的出現頻率 新增影片與頻道重複降權機制，降低首頁長時間反覆出現相同內容的情況 加入基礎隨機排序與近期點擊紀錄，使首頁推薦能在 YouTube 原本的候選內容中產生更多變化 加入 ON / OFF 開關與本機設定儲存 所有資料皆保存在瀏覽器本機，不使用外部伺服器 English Released the initial version of YouTube Randomizer Added local homepage recommendation re-ranking and tracking for recently displayed videos and channels Introduced duplicate penalties to reduce repeated videos and repeated creators Added basic randomization and recent click tracking to increase variation within YouTube's existing candidate pool Added an ON / OFF toggle with locally saved settings All data remains stored locally in the browser with no external server 第二次更新 / Second Update — V0.6.0 中文 大幅更新推薦演算法，開始參考近期觀看紀錄、實際觀看時間、觀看比例、頻道偏好與 YouTube 原始推薦排序，使新內容探索更加自然，同時降低完全不相關影片出現的機率 新增對 Shorts、倍速播放、背景播放與不同影片來源的行為判斷 改善多興趣、多語言、重播型與頻道偏好型使用者的推薦表現 修正：無限捲動異常、多分頁同步問題、Unknown 頻道誤判、一般關鍵字造成的興趣誤判、歷史紀錄競爭問題、擴充功能更新後舊分頁錯誤 所有學習資料仍只保存在瀏覽器本機，沒有加入定位功能或外部資料傳輸 English Major update to the recommendation algorithm. Randomizer now considers recent watch history, actual watch duration, watch ratio, channel preferences, and YouTube's original recommendation order, making discovery more natural while reducing completely unrelated recommendations Added behavior detection for Shorts, playback speed, background playback, and different video sources Improved recommendation behavior for users with multiple interests, multiple languages, repeat-viewing habits, and strong channel preferences Fixed: infinite scroll issues, multi-tab synchronization problems, Unknown channel misclassification, interest misclassification caused by generic keywords, history record conflicts, errors in old tabs after extension updates All learning data remains stored locally in the browser. No location tracking or external data transmission was added 第三次更新 / Third Update — V0.9.0 中文 重新設計 Refresh Exploration 擴張演算法：原本的線性擴張方式改為階段式探索機制，讓每次重新整理有更明確的探索方向 探索現在會逐步經過：新影片 → 新頻道 → 相鄰主題 → 更廣泛的新主題 連續找不到想看的內容時，會逐步降低本輪已反覆出現的影片、頻道與主題權重，並提高尚未出現過內容的曝光比例 新增 Stage-aware Exploration，依照 Refresh Streak 調整不同階段：Baseline、Video Rotation、Topic Probe、Adjacent Expansion、Broad Discovery、Deep Discovery 主題探索現在會更早開始，並在高探索階段更積極測試新的興趣方向 新增 Session Anti-Repeat，使同一輪已經看過多次的影片與頻道更難再次進入首頁前排 Deep Buffer 改為依據目前缺少的新影片、新頻道與主題動態決定是否繼續取得候選內容，降低不必要的等待時間 改善有效觀看判斷：單純點擊影片不會立即重置探索，必須實際觀看一段時間後才會確認找到有效內容 修正 F5、YouTube Logo 與首頁按鈕可能重複增加 Refresh Streak 的問題，現在一次操作只會計算一次 新增繁體中文與英文教學介面 重新設計多種使用者興趣結構模擬，包括單一興趣、多元興趣、簡單與複雜興趣、稀疏資料、興趣漂移與長尾使用者，並加入大量隨機模擬驗證 English Redesigned the Refresh Exploration algorithm: the previous mostly linear expansion system has been replaced by a stage-based discovery model, giving each refresh a clearer exploration purpose Discovery now gradually expands through: New videos → New channels → Adjacent topics → Broader new topics When repeated refreshes fail to find meaningful content, Randomizer gradually reduces the ranking weight of videos, channels, and topics already shown during the current session while increasing exposure to unseen content Added Stage-aware Exploration with multiple refresh stages: Baseline, Video Rotation, Topic Probe, Adjacent Expansion, Broad Discovery, Deep Discovery Topic exploration now begins earlier and becomes significantly more aggressive during higher exploration stages Added Session Anti-Repeat to strongly reduce videos and channels repeatedly shown during the same exploration session Deep Buffer now decides whether to retrieve additional candidates based on missing video, channel, and topic diversity instead of relying only on a fixed candidate target, reducing unnecessary waiting time Improved meaningful-watch detection: simply clicking a video does not immediately reset exploration — it must actually be watched for a period of time before the session is considered successful Fixed duplicate Refresh Streak increments caused by F5, the YouTube logo, and homepage navigation — each user action is now counted only once Added Traditional Chinese and English onboarding Expanded testing to cover multiple user-interest structures, including single-interest, multi-interest, simple and complex interests, sparse profiles, interest drift, and long-tail behavior, along with large-scale randomized simulations