Datathon FME · Smadex Challenge / 2025
Purchase Intent Prediction.
Signals of spending in user behavior
The Smadex challenge asked our team to find predictive signals in high-dimensional user, device and advertising data. Although the challenge is described as purchase-intent prediction, the target in the public notebooks is seven-day in-app purchase revenue, rather than a binary buyer label.
LightGBMDask / ParquetCategorical embeddingsFT-TransformerTemporal validation
From raw events to useful features
Categorical advertiser and device attributes, numerical activity summaries and missing values all need consistent preprocessing. The notebooks explore large Parquet datasets, transformations, categorical mappings and validation splits under the limited time available in a datathon.
Comparing practical models
The code includes a LightGBM regression baseline and neural models with learned embeddings for categorical variables, including an MLP and an FT-Transformer experiment. These approaches offer different trade-offs between feature engineering, training complexity and handling sparse categories.
Result and limits
Our team won the Smadex challenge at Datathon FME 2025. The public repository documents the modeling approaches, but does not contain a verified final leaderboard score. The award was a team result.
Collaborative hackathon entry. The public notebooks support the seven-day purchase-revenue task and methods; the prize is reported in the portfolio CV, without a public score table.