Have a credit risk modelling interview in 20 min.

Very helpful, thank you.

Great stuff! A timely and interesting subject, presented clearly - great tempo - and clear, practical advice.

Thank you so much (drotsiherbalcure) for the difference you make in the lives of your patients! Your kindness, sincere caring, treatment and concern make everything better and you are a great encouragement to humanity. Keep saving lives Dr. thanks for curing my Herpes virus.

Tha k you lot of insight and practical interpretation.

Nice presentation. Please use Python's faker library to produce fake data in real time.

dataset please

bookmarking 23:08

Without the data, I am unable to see any estimands. For all the 3 estimands it shows no such variables found. How can I know what variables to adjust ?

Very useful, thank you!

This is a highly informative and useful presentation. It is clear, concise, and to the point.

How would you get around this problem - you have 2 sets of results from 2 different search engines - for example, one is vector and the other one is full-text. However, it just so happens that the vector search results are super good but the full-text search results are really crappy for this particular query (Not always). Now you apply the Reciprocial Rank Fusion algorithm and it blends together crap and quality instead of keeping more quality and discarding more crap. Wish there was a way to address this problem other than Elastic's custom "script_score" which is basically a static function and assumes that the same scoring algorithm will be applied regardless of input (results)

Excellent presentation. Thank you.

這些是有關資料科學的時間戳記和摘要： 00:47 🧐 這個演示重點在於資料科學中的高級特徵工程，強調策略而不是代碼。 02:08 📈 在分析、大數據和人工智慧方面的投資回報率非常高，引發了對這一領域的興趣。 03:20 📉 據Gartner的炒作週期，目前人工智慧和資料科學領域正處於"幻滅的低谷"，這表示存在挑戰。 05:39 📊 據報告，資料科學專案達到生產的失敗率高達87％，突顯了一個重大問題。 07:42 🧮 有效的特徵工程對於實現資料科學專案的價值和克服挑戰至關重要。 10:59 🌟 選擇和處理特徵通常比選擇算法更重要。 14:10 🎯 特徵工程的主要目標包括處理異常值、缺失值、縮放、降維、平滑化以及處理空值或零值。 16:13 🧩 在特徵工程中，簡化對於降低模型的複雜性和潛在故障點非常重要。 19:30 🔄 對於單個變數的不同轉換的自動測試可以簡化特徵工程過程。 21:10 🤖 考慮將每個單獨的變數視為潛在模型，探索不同的特徵工程轉換。 22:29 🎯 在特徵工程中，可以使用目標均值編碼來處理分類變數，用目標變數均值替代類別。 23:09 📊 決策樹可以幫助確定連續變數的最佳分組邊界，改善分組過程。 24:47 🌲 隨機森林可以自動生成特徵重要性列表，有助於降維，也可以視為一組決策樹或特徵工程。 24:59 🌍 位置數據，如緯度和經度，可以是建模中至關重要的特徵，即使它們看似相近。轉換套件可以幫助有效處理這些數據。 25:27 📅 日期對於特徵工程非常有價值，特別是用於理解購買行為。考慮因素包括一周的哪一天、節假日和地區差異。 26:50 💎 有時，二進制編碼（零或一）可以比連續變數更強大，特別是用於檢測特定事件，如高端商店的購買。 27:31 📊 決策樹可用於為直方圖創建分組，每個葉節點代表一個不同的類別。 28:43 🏛 資料科學家和實施團隊之間的合作對於降低資料科學專案的失敗率至關重要。 29:24 🌳 從決策樹開始可以使模型開發更加容易理解，可以更好地與利益相關者溝通。 30:33 📈 無監督學習可以幫助您探索數據並改進特徵工程，特別是在處理時間序列數據時。 32:13 📉 在編碼稀疏、高維、逗號分隔的特徵，如URL訪問歷史時，考慮按層次分組以獲得更好的結果。 34:28 📊 決策樹可用於確定連續變數轉換為分類變數的分組限制。 35:21 🗳 通過對多數類別進行欠取樣來解決類別不平衡問題，以平衡數據並提高模型性能。 50:23 📊 在處理缺失值時，考慮使用不同的模型來處理帶有缺失值的記錄和完整數據的記錄，或者使用一種基於樹的方法，為空值單獨創建一個類別。 52:29 📊 處理許多分類值，特別是高基數的情況，對於大多數算法可能不是一個重大問題，因為在特徵選擇過程中，許多這些值可能會被消除。 這些時間戳記提供了有關資料科學和特徵工程的寶貴信息，強調了策略、挑戰以及有效處理數據的重要性。特別是在處理缺失值和分類變數時，有一些實用的方法和建議可以幫助您在資料科學專案中取得成功。如果您有任何更具體的問題或需要進一步的解釋，請隨時提問。

Whats funny looking back to this now is that moment google stepped back? That was when they first got BERT to a pre-RLHF GPT-3 level of competence, but the rumor is some execs got spooked and backburnered it. And 2.5ish years on people started unironically intentionally using bing for the first time since they downloaded chrome. I expect those execs got canned but i havent followed closely.

that bias issue is pretty interesting on how to know when to retrain the data, I guess the model could be deployed on about 40-50 percent of the current users who are newly signing up. Then do an iteration of an AB test again on those results, and similarly in a loop until a certain threshold has been reached where mostly everyone is getting a sign up post calls. It would still not be 100% as nothing can be 😅 but might provide more data for furrher iterations of the model with less bias than if we just deployed it on all users together and then try to retrain it.

Awesome presentation. Can anyone suggest resources (books, courses) to upskill in data science, specifically in the marketing related domain?

please share slide ?

Nice sum up of these packages for feature engineering

Such underrated video

Thank you Jacqui for this useful and well delivered presentation 😊

Excelent thought: "Visualizations will act as a campfire around which everyone will gather to tell stories."

Audio sucks

You are my inspiration, Lisa! I have been getting so much following your pathways and your learning resources are extremely helpful. Your dedication and commitment to learning inspired me as well. Keep up the magnificent work in data, and may your journey be filled with continued growth and success!

After following this presentation, I took 3 minutes to run stacknet and it moved me up about 25 positions on the leaderboard. By doing feature engeneering & selection I think performance will increase even more.

@Data Science Festival where is the link to the pack referenced in the talk, please?

Can you prepare for restaurants options if they are starting straight away prepare food or if they are busy at their own restaurant they will start to work on delivery food after 5min, 10min or 15 min.

Excellent, thank you

What an impeccable presentation! As an expert from deep tech myself, this is simply the best talk among all the sessions I attended on that DSF day.

It would be great to have the speaker's name and affiliation in the description.

Very useful!

If the tutorial could have time stamps

Thank you for providing such informative insights. Undoubtedly, the utilization of graph databases is essential for professionals working in the field of data.

Excellent. Thank you Nichola. Interesting to hear the practical issues you are encountering and solving.

Great stuff! Timely, interesting and clear. - Thank you Kris. I experimented with ggerganov/whisper.cpp. In case it helps anyone ... Audacity wav exports have to be signed 16 bit pcm to work. Wav files are quite large, even for small samples. Congratulations on the promotion :)

Very interesting!. Old project. Give Feedback. Creation of learning plan. Why rather tan what?. Keeping it relevant. Protect time for learning. Create passion as spark joy.

Thank you for this great video, very informative.

I kind of disagree about call_item_item() function. Because we search similar items of same kind (candidate and candidate) instead of using dot product we need to use something like tf.keras.losses.cosine_similarity() to find nearest neighbors. For functions like call_user_items() or call_item_users() we can use tf.keras.layers.Dot() because those are query and candidate items.

terrible indian english 🤮

i really like your project. I want to understand clearly about algorithm. Can you share me some documents?

where is thegithub link..??

thanks. how would i apply L1/L2 to resolve overfitting problem?

1:22:21 - Some users have more than 1 product. By randomly generating numbers do not guarantee that the products are not owned by an owner. For an example, user 'PIXcm7Ru5KmntCy0yA1K' has 3 products namely [10524048, 9870070, 11574730]. Random generation of indices can unfortunately end up being 9870070. You can code such that these 10 indices do not end up in the situation.

The Github repo containing the code from the talk is here: github.com/Tommo565/dask-presentation The Slides used in the talk are here: docs.google.com/presentation/d/1YBVjHGwTIZvyor8E6UUT7gMCymbZdrcj0RY2cy7zRvY/edit?usp=sharing The Dask site is here: www.dask.org/

Great presentation 👍🏾👍🏾👍🏾

Excellent presentation by Ailish. She explained the various measures very clearly with detailed examples. Usually, no one likes to explain such concepts so clearly with examples but these so called 'experts' talk round and round using jargons. However, it is amusing to watch her discomfort when answering questions!! In fact, the questions were very easy to answer compared to what she explained.

Wow such an informative video ! 💯🙌🏻

It helps a lot on my case study of Zopa ! REALLY APPRECIATE 🤩

Great presentation! Can the notebook + data be found somewhere?

Very interesting presentation. I am inspired!