README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - text-classification
tags:
  - TAAC2026
  - recommendation
size_categories:
  - n<1K

TAAC2026 Demo Dataset

⚠️ Update[2026.04.10]: This demo dataset has been updated to newest version with the following changes:

• The parquet file is now a flat column layout, with all features as top-level columns.
• Add a sequence feature and update some user/item features. Participants should refer to the updated demo_1000.parquet and this README for the latest schema and data details.

A sample dataset containing 1,000 user-item interaction records for the TAAC2026 competition. This dataset uses a flat column layout — all features are stored as individual top-level columns instead of nested structs/arrays.

Dataset Overview

Property	Value
File	demo_1000.parquet
Rows	1,000
Columns	120
Format	Apache Parquet
File Size	~38.38 MB

Label Distribution

label_type	Count	Percentage
1	876	87.6%
2	124	12.4%

Column Categories

The 120 columns fall into 6 categories:

Category	Count	Arrow Type	Description
ID & Label	5	int64 / int32	Core identifiers, label, and timestamp
User Int Features	46	int64 / list<int64>	Integer-valued user features (scalar or array)
User Dense Features	10	list<float>	Float-array user features
Item Int Features	14	int64 / list<int64>	Integer-valued item features (scalar or array)
Domain Sequence Features	45	list<int64>	Behavioral sequence features from 4 domains

---

Detailed Column Schema

ID & Label Columns (5 columns)

Column	Arrow Type	Nulls	Min	Max	Mean	Unique
user_id	int64	0	2,727,076	12,728,427	7,835,799.34	1,000
item_id	int64	0	6,854	278,202,253	112,417,687.39	837
label_type	int32	0	1	2	1.124	2
label_time	int64	0	1,772,725,027	1,772,725,910	1,772,725,503.90	553
timestamp	int64	0	1,772,725,000	1,772,725,781	1,772,725,275.45	501

User Int Features (46 columns)

Scalar Columns (int64)

Column	Nulls	Null%	Min	Max	Mean	Unique
user_int_feats_1	0	0.0%	1	4	3.381	3
user_int_feats_3	30	3.0%	9	1,839	987.557	341
user_int_feats_4	30	3.0%	1	986	498.813	268
user_int_feats_48	2	0.2%	3	99	58.006	52
user_int_feats_49	7	0.7%	1	2	1.582	2
user_int_feats_50	4	0.4%	0	1	0.998	2
user_int_feats_51	1	0.1%	40	150	56.157	5
user_int_feats_52	1	0.1%	5	174	93.856	36
user_int_feats_53	1	0.1%	3	557	288.542	264
user_int_feats_54	368	36.8%	3	2,843	1,476.783	462
user_int_feats_55	19	1.9%	8	41	29.682	13
user_int_feats_56	19	1.9%	1	1,434	752.658	405
user_int_feats_57	31	3.1%	2	250	126.588	105
user_int_feats_58	150	15.0%	1	2	1.699	2
user_int_feats_59	150	15.0%	1	14	8.371	8
user_int_feats_82	204	20.4%	1	23	9.097	23
user_int_feats_86	692	69.2%	2	245	105.474	61
user_int_feats_92	494	49.4%	1	2	1.500	2
user_int_feats_93	171	17.1%	1	37	14.667	36
user_int_feats_94	521	52.1%	1	6	3.770	6
user_int_feats_95	318	31.8%	1	3	2.758	3
user_int_feats_96	678	67.8%	1	3	1.817	3
user_int_feats_97	292	29.2%	1	3	1.599	3
user_int_feats_98	103	10.3%	1	3	2.678	3
user_int_feats_99	812	81.2%	1	3	2.936	2
user_int_feats_100	845	84.5%	1	2	1.955	2
user_int_feats_101	910	91.0%	2	3	2.956	2
user_int_feats_102	877	87.7%	1	3	1.130	2
user_int_feats_103	862	86.2%	1	3	2.717	3
user_int_feats_104	372	37.2%	1	3	2.360	3
user_int_feats_105	309	30.9%	1	3	2.287	3
user_int_feats_106	160	16.0%	1	3	1.760	3
user_int_feats_107	300	30.0%	1	2	1.094	2
user_int_feats_108	516	51.6%	2	7	5.455	6
user_int_feats_109	854	85.4%	1	7	2.993	7

Array Columns (list<int64>)

Column	Nulls	Null%	Element Type
user_int_feats_15	139	13.9%	list<int64>
user_int_feats_60	592	59.2%	list<int64>
user_int_feats_62	70	7.0%	list<int64>
user_int_feats_63	70	7.0%	list<int64>
user_int_feats_64	70	7.0%	list<int64>
user_int_feats_65	80	8.0%	list<int64>
user_int_feats_66	86	8.6%	list<int64>
user_int_feats_80	200	20.0%	list<int64>
user_int_feats_89	55	5.5%	list<int64>
user_int_feats_90	91	9.1%	list<int64>
user_int_feats_91	450	45.0%	list<int64>

User Dense Features (10 columns)

All columns are list<float> arrays (e.g. embedding vectors).

Column	Nulls	Null%	Description
user_dense_feats_61	2	0.2%	256-dim embedding vector
user_dense_feats_62	70	7.0%	Variable-length float array
user_dense_feats_63	70	7.0%	Variable-length float array
user_dense_feats_64	70	7.0%	Variable-length float array
user_dense_feats_65	80	8.0%	Variable-length float array
user_dense_feats_66	86	8.6%	Variable-length float array
user_dense_feats_87	15	1.5%	320-dim embedding vector
user_dense_feats_89	55	5.5%	Variable-length float array
user_dense_feats_90	91	9.1%	Variable-length float array
user_dense_feats_91	450	45.0%	Variable-length float array

Item Int Features (14 columns)

Column	Arrow Type	Nulls	Null%	Min	Max	Mean	Unique
item_int_feats_5	int64	2	0.2%	4	325	118.452	82
item_int_feats_6	int64	2	0.2%	0	977	419.073	216
item_int_feats_7	int64	2	0.2%	0	2,806	1,052.866	349
item_int_feats_8	int64	2	0.2%	-1	2,431	463.712	226
item_int_feats_9	int64	2	0.2%	3	37	21.171	24
item_int_feats_10	int64	2	0.2%	2	309	150.007	110
item_int_feats_11	list<int64>	439	43.9%	—	—	—	—
item_int_feats_12	int64	2	0.2%	0	2,777	1,039.381	352
item_int_feats_13	int64	2	0.2%	1	8	4.457	8
item_int_feats_16	int64	2	0.2%	2	35,259	12,356.101	662
item_int_feats_81	int64	2	0.2%	0	2	0.508	3
item_int_feats_83	int64	832	83.2%	1	31	17.595	22
item_int_feats_84	int64	832	83.2%	3	226	131.131	66
item_int_feats_85	int64	832	83.2%	4	1,001	439.816	103

Domain Sequence Features (45 columns)

Variable-length list<int64> sequences from 4 behavioral domains:

Domain	Columns	Count	Nulls per Col	Max Seq Length
domain_a	domain_a_seq_38 – _46	9	5	1,888
domain_b	domain_b_seq_67 – _79, _88	14	12	1,952
domain_c	domain_c_seq_27 – _37, _47	12	2	3,894
domain_d	domain_d_seq_17 – _26	10	80	3,951

---

Null Coverage Summary

Group	Columns	Zero Coverage	Low Coverage (<50%)	Notes
user_int_feats_	46	0	11	Columns 99–103, 109 have >80% nulls
user_dense_feats_	10	0	0	user_dense_feats_91 has 45% nulls
item_int_feats_	14	0	3	item_int_feats_83–85 have ~83% nulls
domain_a_seq_	9	0	0	Very low null rate (0.5%)
domain_b_seq_	14	0	0	Low null rate (1.2%)
domain_c_seq_	12	0	0	Very low null rate (0.2%)
domain_d_seq_	10	0	0	Moderate null rate (8.0%)

High-Null Columns (>50% null)

Column	Null Count	Null%
user_int_feats_101	910	91.0%
user_int_feats_102	877	87.7%
user_int_feats_103	862	86.2%
user_int_feats_109	854	85.4%
user_int_feats_100	845	84.5%
item_int_feats_83	832	83.2%
item_int_feats_84	832	83.2%
item_int_feats_85	832	83.2%
user_int_feats_99	812	81.2%
user_int_feats_86	692	69.2%
user_int_feats_96	678	67.8%
user_int_feats_60	592	59.2%
user_int_feats_94	521	52.1%
user_int_feats_108	516	51.6%
user_int_feats_92	494	49.4%
user_dense_feats_91	450	45.0%
user_int_feats_91	450	45.0%
item_int_feats_11	439	43.9%

---

Usage

import pyarrow.parquet as pq
import pandas as pd

# Read the parquet file
pf = pq.ParquetFile("data_1000/demo_1000.parquet")
table = pf.read()
df = table.to_pandas()

print(df.shape)       # (1000, 120)
print(df.columns)     # ['user_id', 'item_id', 'label_type', ...]

# Check label distribution
print(df['label_type'].value_counts())
# 1    876
# 2    124

# Access a sequence feature
seq = df['domain_a_seq_38'].dropna().iloc[0]
print(type(seq), len(seq))  # <class 'numpy.ndarray'> variable length

# Access an embedding feature
emb = df['user_dense_feats_61'].dropna().iloc[0]
print(type(emb), len(emb))  # <class 'numpy.ndarray'> 256

# Null analysis
null_pct = df.isnull().mean().sort_values(ascending=False)
print(null_pct[null_pct > 0.5])  # Columns with >50% nulls

Relationship to Other Files

File	Rows	Size	Compression	Description
data_1000/demo_1000.parquet	1,000	~38 MB	None	This dataset — first 1,000 rows
demo_data/demo_1000_0408.gz.parquet	1,016	~27 MB	Gzip	Full 1,016-row source dataset
test_demo_data/sample_10.parquet	10	~548 KB	—	10-row test sample

Key Notes

1. Nullable int64: All *_int_feats_* scalar columns are stored as Arrow int64 with native null support. When reading with pandas, nullable int columns may be converted to float64 — use df[col].fillna(-1).astype(int) or read with pd.Int64Dtype() to preserve the integer type.
2. No nested structs: Unlike the older sample_data.parquet, all features are flat top-level columns.
3. Sparse features: 18 columns have >40% null values — handle missing data carefully during feature engineering.
4. Sequence lengths vary widely: Domain sequences range from length 1 to ~3,951, which may require truncation or padding for model input.
5. Imbalanced labels: ~87.6% label_type=1 vs ~12.4% label_type=2 — consider class balancing strategies.