README.md

metadata

license: cc-by-nc-4.0
tags:
  - tsfile
  - time-series
  - recommendation
  - taac2026
modality: timeseries
pretty_name: TAAC2026 Data Sample 1000 TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.tsfile

TAAC2026 Data Sample 1000 TsFile

This dataset is a TsFile conversion of TAAC2026/data_sample_1000, the TAAC2026 demo recommendation dataset with 1,000 user-item interaction records.

Modalities: Time-series. The source dataset is a flat Parquet file where all features are top-level columns. Some top-level columns are variable-length list features, so the conversion stores scalar interaction features and sequence features in separate TsFile tables instead of expanding the lists into more than 130,000 wide columns.

Source Dataset

• Original dataset: TAAC2026/data_sample_1000
• Source file: demo_1000.parquet
• Source license: cc-by-nc-4.0
• Source tags: TAAC2026, recommendation
• Source scale: 1,000 rows, 120 top-level columns, about 39 MB
• Source schema groups: 5 ID/label columns, 46 user integer features, 10 user dense features, 14 item integer features, and 45 domain sequence features

Converted Files

The upload contains 9 TsFile files under data/.

File	Table	Rows	Notes
data/data_sample_1000_scalar.tsfile	data_sample_1000_scalar	1,000	Scalar interaction features
data/data_sample_1000_user_int_lists.tsfile	data_sample_1000_user_int_lists	11,560	11 user integer list features
data/data_sample_1000_user_dense_lists.tsfile	data_sample_1000_user_dense_lists	318,538	10 user dense list features
data/data_sample_1000_item_lists.tsfile	data_sample_1000_item_lists	2,086	1 item list feature
data/data_sample_1000_domain_a_seq.tsfile	data_sample_1000_domain_a_seq	701,086	9 domain A sequence features
data/data_sample_1000_domain_b_seq.tsfile	data_sample_1000_domain_b_seq	570,758	14 domain B sequence features
data/data_sample_1000_domain_c_seq.tsfile	data_sample_1000_domain_c_seq	449,431	12 domain C sequence features
data/data_sample_1000_domain_d_seq_1.tsfile	data_sample_1000_domain_d_seq	1,048,576	Domain D sequence shard 1
data/data_sample_1000_domain_d_seq_2.tsfile	data_sample_1000_domain_d_seq	51,283	Domain D sequence shard 2

The sequence tables contain 3,153,318 rows in total. domain_d_seq was split into two TsFile shards by the TsFile conversion tool.

Schema Design

Scalar event table:

• Time: epoch milliseconds derived from source timestamp * 1000
• TAG columns: event_index, user_id
• FIELD columns: item_id, label_type, label_time, event_timestamp, and all source scalar user/item features

Sequence tables:

• Time: sequence_index within each source interaction event
• TAG columns: event_index, user_id, item_id
• FIELD columns: event_timestamp, label_time, label_type, plus the list feature values for that sequence family
• event_timestamp preserves the source timestamp value in seconds

Conversion Notes

• No source columns are intentionally dropped.
• Source timestamp is renamed to event_timestamp and also used to create the scalar event table's TsFile Time.
• Variable-length list columns are reshaped into per-family sequence tables. This preserves list positions while avoiding an extremely wide table.
• Missing list positions caused by unequal sequence lengths are stored as nulls.
• The converted layout keeps one event_index per original source row so users can join scalar and sequence tables back to the original interaction record.

Validation

Local validation confirmed that all 9 TsFile files are non-empty. TsFile metadata row counts match the staged Parquet row counts for every table, including both shards of data_sample_1000_domain_d_seq.

Minimal Read Example

from tsfile import TsFileReader

path = "data/data_sample_1000_scalar.tsfile"
with TsFileReader(path) as reader:
    schemas = reader.get_all_table_schemas()
    print(schemas.keys())