asaander719/coupledhsr

This is a research code repository for TacoRec, an AI model designed to study how people behave on e-commerce platforms and predict what they might do next. Specifically, it looks at multiple types of user actions at once, such as clicking, adding to cart, favoriting, and purchasing, rather than treating each behavior in isolation. The goal is to make more accurate recommendations by understanding the connections between these different actions. The repository works with two datasets drawn from real e-commerce platforms: one from JD (a Chinese online retailer) and one called UB (UserBehavior). Both datasets contain hundreds of thousands of user interaction sequences. Scripts in the repo convert these datasets into a format that works with RecBole, a research toolkit for building and testing recommendation models. To use the code, a researcher would first run a conversion script to prepare the dataset files, then run a training script specifying which model and dataset to use. The README provides separate commands for running on a single behavior type (for example, just purchase data) or across all behavior types at once. There is also a hyperparameter sweep script that tries different combinations of settings, such as learning rate and model size, and records results including accuracy, training time, GPU memory usage, and computational cost. The repository also includes commands for running a baseline model called AMPL, which lets researchers compare TacoRec's results against an existing method. An important note in the README explains that the data conversion script does not apply the original dataset's preprocessing filters itself, it assumes those filters have already been applied and only handles the format conversion step. The README is written primarily for researchers reproducing or building on this work. There is no explanation of the model's internal design or links to a paper.