Overview

🏆

Welcome to the Music Recommendation System Challenge! 🎶

Our Team:

• Djalil — Product Owner at hitter & Izi, wants every user to discover the perfect soundtrack for their day 🎧
• Darkhan — Data Scientist, tasked with turning mountains of user data into smart, personalized recommendations 💡
• Adilkhan — an active user of the platform, curious and eager for new music discoveries 🎵

Your mission: help Darkhan build a recommendation system that predicts the top 50 most relevant tracks for each user based on their listening history and preferences — making sure users like Adilkhan get the songs they love most.

The dataset comes from real interactions on hitter and Izi, where users listen, like, and download tracks daily. You can explore any approach — from classic machine learning to deep learning or hybrid models.
Think of it as crafting the ultimate playlist that keeps Adilkhan hitting “play” again and again! 🚀

---

🎯 Goal

Build a model that recommends 50 tracks per user, aiming to predict which songs each listener will enjoy the most — and how much they’ll actually listen.

Help Darkhan impress Djalil with a model that truly understands users like Adilkhan, turning raw data into a seamless, personalized music experience. 🔥

Evaluation

📊 Evaluating Recommendation Quality

Here’s how we measure who made the best recommendations:

1. Predict Tracks
   For each user (e.g., Adilkhan!), your model predicts a list of 50 tracks they might like.

2. Check Listening Fractions
   For each recommended track, we record how much the user actually listened:

   • 0.0 — not listened at all
   • 0.25 — listened a quarter of the track
   • 0.5 — listened halfway
   • 0.75 — listened three-quarters
   • 1.0 — listened fully (replays don’t count extra)

3. User Score
   Sum these fractions across the 50 tracks → this gives a score for that user (range: 0 to 50).

4. Average per User
   Compute the mean of all users’ scores → tells us, on average, how many tracks each user actually listened to (range : 0 to 50).

5. Final Score
   normalized score between 0 and 1.

📈 How to Interpret the Metric

• Score ≥ 0.25 → Users are starting to engage with your recommendations. Even a quarter of a track counts as a small win!
• Score ~0.5 → Users listen to about half of your recommendations. Nice job! Your suggestions are clearly interesting.
• Score ~0.75 → Most tracks are listened to three-quarters of the way. Great work! Users really like your recommendations.
• Score ~1.0 → Users fully listen to almost all recommended tracks. Excellent! You’re nailing it.

Submission

Overview

Your submission should contain ranked recommendations for each user in the test set. Each user must have exactly 50 item recommendations ranked from 1 to 50, where rank 1 is the most recommended item and rank 50 is the least recommended.

File Format

• File Type: CSV file

Column Specifications

Column	Type	Description	Constraints
id	Integer	Unique row identifier	Sequential starting from 0
user_id	Integer	User identifier from test set	Must match test set user_ids
item_id	Integer	Item identifier to recommend	Must be valid item_id
rank	Integer	Recommendation rank	1-50 (1 = highest, 50 = lowest)

Requirements

• All columns are mandatory - missing any column will result in submission error
• Each user in the test set must have exactly 50 recommendations
• Ranks must be unique per user-item (1 through 50)
• No duplicate item recommendations for the same user
• Colum id is mandatory

Example , how to create id column:

recos.insert(0, 'id', range(len(recos)))

Sample Submission

id	user_id	item_id	rank
0	1	123	1
1	1	3123	2
2	1	3213123	3
3	1	123312312	4
4	1	112	5
5	1	1	6
6	1	73	7
7	1	99	8
...	...	...	...
49	1	705	50

Dataset Description

This dataset is derived from real user interactions on the hitter and Izi music platforms. It includes user-track interactions, track metadata, and user metadata.

The training set covers production-level user behavior from 2025-02-28 to 2025-08-30 (a 6-month period).
The test set represents user interactions from 2025-09-01 to 2025-09-15, used to evaluate recommendation performance. It only includes users who appear in the training interactions (no cold-start users).

⚠️ Note: interactions.csv may contain noise, duplicates, or minor bugs reflective of real-world data.
⚠️ Note: item_metadata.csv, user_metadata.csv may have missing or incomplete records.

---

Files

• interactions.csv — Records of user interactions with tracks
• item_metadata.csv — Metadata about tracks
• user_metadata.csv — Metadata about users
• test.csv — List of users for whom predictions must be made

---

interactions.csv

Column	Description	Data Type	Notes
user_id	Unique identifier for the user.	Integer	References users in user_metadata.csv.
item_id	Unique identifier for the track (item).	Integer	References tracks in item_metadata.csv.
listened_duration	Duration (in seconds) the user listened to the track in this interaction.	Integer	May be partial or noisy; includes skips/outliers.
listened_datetime	Timestamp of when the interaction occurred.	Datetime	Format: YYYY-MM-DD HH:MM:SS.ssssss.

🕒 Time Range: 2025-02-28 → 2025-08-30 (6 months of production-level data)

---

item_metadata.csv

Column	Description	Data Type	Notes
item_id	Unique identifier for the track.	Integer	Primary key for tracks.
track_name	Name/title of the track.	String	May include special characters or non-English text.
artist_name	Name of the artist(s) or performer(s).	String	Can include collaborations (e.g., “Artist feat. Other”).
track_duration	Full duration of the track in seconds.	Integer	Total length of track.
track_genres_list	Comma-separated list of genres associated with the track.	String	Example: ['Genre1', 'Genre2']; may include non-English genres.
track_dislike_count	Total dislikes received by the track.	Integer	Aggregate metric; starts at 0 for new tracks.
track_like_count	Total likes received by the track.	Integer	Aggregate metric; starts at 0 for new tracks.
track_download_count	Total downloads of the track.	Integer	Aggregate metric; starts at 0 for new tracks.

---

user_metadata.csv

Column	Description	Data Type	Notes
user_id	Unique identifier for the user.	Integer	Primary key for users.
age_bin	Binned age group of the user (based on the ML model).	String	Examples: '18_29', '30_44', '45_60', '61_inf'.
children	Number of children the user has (based on the ML model).	Integer	0–1
gender	Gender of the user (based on the ML model).	String	'M' for male, 'F' for female.
top_genre	User's preferred or inferred top genre.	String	Examples: 'Поп', 'Джаз', 'Chill'; may be in non-English.
user_disliked_track_count	Total number of tracks disliked by the user.	Integer	Aggregate user-level metric.
user_liked_track_count	Total number of tracks liked by the user.	Integer	Aggregate user-level metric.
user_downloaded_track_count	Total number of tracks downloaded by the user.	Integer	Aggregate user-level metric.

---

test.csv

Column	Description	Data Type	Notes
user_id	Unique identifier for the user.	Integer	All users exist in the training interactions; no cold-start users.

🧭 Time Range: 2025-09-01 → 2025-09-15 (interactions period for evaluation)

---

Prediction Format

The submission file must contain recommendations for each user in the test set, ranked by relevance or preference.

Column	Description	Data Type
id	Sequence of id .	Integer
user_id	User identifier (matches test.csv).	Integer
item_id	Recommended track identifier.	Integer
rank	Rank of the recommendation (starting from 1 to 50).	Integer

kaggle competitions download -c hear-me-personalized-music-recommender