Trivago Recsys Experiments

How to use:

1. These scripts expect a data/trivago subdirectory...

   • mkdir data
   • put the trivago dataset in data/ OR if using ada, create a link to the trivago folder in the storage node.
   • the files you need are: train.csv, validation.csv, confirmation.csv

2. compute features for the experiments:

   • python3 load_trivago.py
   • python3 hash_user_id.py
   • python3 load_trivago_blind.py

3. run the experiments:

   • python3 train_eval.py
   • python3 train_eval_blind.py

More details on each script:

load_trivago.py

This script creates a readable parquet file which contains all the data which is required to train the models. In this file, I've included a range of dataclasses to help me understand the dataset, and to make feature extraction more intuitive and readable. The objects I've defined are as follows:

1. Hotel
2. Interaction
3. Session
4. UserProfile
5. SessionData

Following LogicAI's 2019 recsys strategy, I computed user features on a rolling basis to prevent overfitting. Particularly, I sorted sessions by their starting timestamp, and added each interaction to a user profile and graph only after features had been extracted from that interaction.

compute_ctr.py

This script computes the click-through ratios (CTRs) for all items.

extract_hotel_features.py

Construct hotel (item-based) features for learning.

hash_user_id.py

We are interested in knowing what happens when we pretend that all users are new visitors to the site. We can't just null out the user based features because we use a graph based model as one of our features, which still may be a useful features even if edges only occur between sessions and items (rather than users and items).

load_trivago_blind.py

This script would remarkably similar to the above, but it excludes user profiles and computes fewer features. It also constructs a user-item graph as a session-item graph instead, which is more sparse.

Future work/stuff I didn't cover this summer.

I tackled a limited range of features which were considered most important in the challenge, while using my best judgement to exclude ones which I considered "cheating". For example, the first-place submission had

Some more features may have been helpful in boosting our final test MRR (.576)

In these experiments, I found that the user was really not important in recommending items over such a short period. That is, people likely do not plan multiple, different vacations/work trips in one week.