VisionX — Financial Alert Intelligence

2Ma'lumotlarDataset overview

Beshta fayl: signallar (train/test), ularning tranzaksiyalari (train/test) va namuna javob fayli. Har bir signalga uning mijozining signalgacha bo'lgan tranzaksiya tarixi biriktirilgan.Five files: alerts (train/test), their transactions (train/test) and a sample submission. Each alert comes with its customer's transaction history up to the alert.

Kelajak ma'lumotidan foydalanmaslik (leakage)No information from the future (leakage)

• Signal sanasida vaqt yo'q, shuning uchun signalni o'sha kunning 00:00 deb oldik. Undan keyingi 3 934 (train) va 1 456 (test) tranzaksiya belgilarda ishlatilmaydi.The alert date has no time, so we treat the alert as 00:00 that day. The 3 934 (train) and 1 456 (test) transactions at or after that moment are never used.
• signal_id modelga berilmaydi (ID tartibi va maqsad: AUC 0.498, bog'liqlik yo'q).signal_id is never a model input (ID order vs target: AUC 0.498, no relation).
• Train va test'ni ajratishga uringan model (adversarial validation) AUC 0.502 oldi: test oddiy tasodifiy tanlanma.A model trying to tell train from test (adversarial validation) scored AUC 0.502: the test set is an ordinary random sample.

3Ma'lumotlar tahlili (EDA)Exploratory data analysis

Olti asosiy grafik; har birining tepasida bir-ikki jumlalik xulosa. Grafiklar interaktiv va faqat yig'ma raqamlardan chizilgan.Six key charts, each with a one or two-sentence takeaway above it. Charts are interactive and drawn from aggregated numbers only.

5Belgilar yaratishFeature engineering

Tranzaksiyalarni qanday qilib bashorat qiluvchi belgilarga aylantirdik? Asosiy g'oya: bitta signalga tegishli yuzlab tranzaksiyalar bitta qatorga umumlashtiriladi.How did we turn transactions into predictive features? The key idea: the hundreds of transactions behind one alert are summarised into a single row.

Yakuniy modelga qaysi belgilar kirdi?Which features does the final model use?

626 ta belgi yaratildi, hammasi modelga beriladi; yakuniy LightGBM ulardan 463 tasini haqiqatda ishlatadi (gain > 0), qolgan 163 tasi hech bir bo'linishda tanlanmagan. Eng muhim 20 belgi umumiy gain'ning 35%, eng muhim 100 tasi 70% qismini beradi (15-rasm). Vaqt naqshlari va pass-through guruhlari CV'ni yaxshilamagani uchun umuman yaratilmaydi.We build 626 features and pass all of them to the model; the final LightGBM actually uses 463 (gain > 0), while the other 163 are never picked for a split. The top 20 features carry 35% of the total gain and the top 100 carry 70% (Fig. 15). Time-pattern and pass-through groups did not improve CV, so they are not built at all.

Har bir belgi guruhini o'chirib, CV qancha tushishini o'lchadik (musbat = guruh foydali). Foyda bermagan guruhlar olib tashlandi.We switched off one feature group at a time and measured the CV drop (positive = the group helps). Groups that did not help were dropped.

6Machine learning modelMachine learning model

• Logistic regression — baseline (oddiy chiziqli model)baseline (a simple linear model)
• CatBoost, XGBoost — gradient boosting nomzodlarigradient-boosting candidates
• LightGBM — gradient boosting nomzodi; validatsiyada eng yaxshisi, yakuniy modelgradient-boosting candidate; best in validation, the final model
• ROC-AUC — baholash ko'rsatkichievaluation metric

Validatsiya qanday qurilganHow validation is set up

• Bo'lish signal darajasida: takroriy stratifikatsiyalangan 5-fold (3 marta, seed 42). Tranzaksiyalar alohida bo'linmaydi; har bir signalning belgilari faqat o'z tranzaksiyalaridan hisoblanadi.The split is at alert level: repeated stratified 5-fold (3 repeats, seed 42). Transactions are never split on their own; each alert's features come only from its own transactions.
• Imputatsiya, masshtablash va erta to'xtatish faqat o'quv fold ichida. Giperparametrlar Optuna bilan topilib, qotirilgan.Imputation, scaling and early stopping only ever see the training fold. Hyper-parameters were tuned with Optuna and then frozen.
• Test sanalari train bilan bir xil oraliqda va adversarial AUC ≈ 0.5, shuning uchun tasodifiy (vaqt bo'yicha emas) bo'lish o'rinli.Test dates cover the same range as train and adversarial AUC ≈ 0.5, so a random (not time-based) split is appropriate.

Qaysi model tanlandi va nima uchunWhich model was chosen, and why

Modellarni aralashtirib ko'rdik (OOF 0.6501). Lekin og'irliklarni 4 foldda topib, 5-foldda tekshirsak, blend 0.6499, eng yaxshi bitta model esa 0.6500. Farq yo'q darajada, shuning uchun sodda variantni tanladik: LightGBM, butun train'da 5 xil seed bilan o'qitilgan.We tried blending the models (OOF 0.6501). But when the weights are fitted on 4 folds and checked on the 5th, the blend gets 0.6499 and the best single model 0.6500. That is no real difference, so we went with the simpler option: LightGBM, trained on all of train with 5 different seeds.

Sayt, CSV va notebook bir-biriga mos. Bu yerda ko'rsatilgan model (LightGBM) yakuniy team_040817EA.csv faylini yaratgan model bilan bir xil (SHA256 96e74eefe3d9…). Notebook xuddi shu faylni qayta yaratadi, va bu avtomatik tekshiriladi (farq ≤ 1e-6).Website, CSV and notebook agree. The model shown here (LightGBM) is the one that produced the final team_040817EA.csv (SHA256 96e74eefe3d9…). The notebook regenerates the same file, and this is checked automatically (difference ≤ 1e-6).

7Model natijalariModel performance