Update dataset card: anti-exploitation filters, 286 sequences

README.md

@@ -7,138 +7,104 @@ task_categories:
 tags:
 - temporal-point-process
 - event-sequences
-- github
-- software-engineering
-- developer-workflows
+- amazon-reviews
 - marked-temporal-point-process
 size_categories:
 - n<1K
-configs:
-- config_name: default
-  data_files:
-  - split: test
-    path: data/test-*
-dataset_info:
-  features:
-  - name: seq_idx
-    dtype: int64
-  - name: seq_len
-    dtype: int64
-  - name: type_category
-    dtype: string
-  - name: span_weeks
-    dtype: float64
-  - name: description
-    dtype: string
-  - name: metadata
-    dtype: string
-  - name: time_since_start
-    list: float64
-  - name: time_since_last_event
-    list: float64
-  - name: type_event
-    list: string
-  - name: type_text
-    list: string
-  splits:
-  - name: test
-    num_bytes: 17358437
-    num_examples: 286
-  download_size: 9705164
-  dataset_size: 17358437
 ---
 
-# GitHub User Event Streams
+# Amazon Product Review Events
 
-Curated event sequences from individual GitHub developers, designed for temporal point process (TPP) and marked temporal point process (MTPP) modeling. Each sequence captures a single developer's activity across multiple repositories over a one-week window.
+Curated per-user product review sequences from [Amazon Reviews 2023](https://huggingface.co/datasets/McAuley-Lab/Amazon-Reviews-2023) (McAuley Lab), designed for temporal point process (TPP) and marked temporal point process (MTPP) modeling. Each sequence captures a single user's chronologically-ordered reviews within one product category.
+
+Sequences are aggressively filtered to prevent pattern exploitation — uninformative "other" events are removed, and sequences with dominant-type or consecutive-repeat shortcuts are dropped to ensure models must perform genuine reasoning.
 
 ## Dataset Description
 
-- **Source:** [GH Archive](https://www.gharchive.org/)
-- **Time window:** June 1 - 7, 2024
-- **Grouping:** Events grouped by user
-- **Sequences:** 382
-- **Sequence length:** 80-100 events per sequence
-- **Time unit:** hours
-- **Language:** English only
+- **Source:** [Amazon Reviews 2023](https://huggingface.co/datasets/McAuley-Lab/Amazon-Reviews-2023) (McAuley Lab, HuggingFace)
+- **Categories:** Electronics, Books, Home & Kitchen, Beauty & Personal Care
+- **Grouping:** Events grouped by user within a single product category
+- **Sequences:** 286
+- **Sequence length:** 80–100 events per sequence (mean: 88.9)
+- **Event types:** 54 sub-categories (no "other")
+- **Time unit:** days
 
 ## Schema
 
-Each record is a dictionary with 8 fields:
+Each record is a dictionary with 10 fields:
 
 | Field | Type | Description |
 |---|---|---|
 | `seq_idx` | int | Sequence index |
 | `seq_len` | int | Number of events in the sequence |
-| `description` | str | User name, active repos (with descriptions), and event window |
-| `metadata` | str (JSON) | `actor_login`, `num_repos`, `top_repos`, `time_unit` |
-| `time_since_start` | list[float] | Time since the first event (in `time_unit`, default hours) |
-| `time_since_last_event` | list[float] | Time since the previous event (in `time_unit`, default hours) |
-| `type_event` | list[str] | Event type labels (see below) |
+| `type_category` | str | Parent product category (e.g., `"Electronics"`) |
+| `span_weeks` | float | Total time span of the sequence (weeks) |
+| `description` | str | Category, user activity window, and review period |
+| `metadata` | str (JSON) | `user_id`, `parent_category`, `num_reviews`, `num_sub_categories`, `date_range_start`, `date_range_end`, `span_weeks`, `time_unit` |
+| `time_since_start` | list[float] | Time since the first event (in days) |
+| `time_since_last_event` | list[float] | Time since the previous event (in days) |
+| `type_event` | list[str] | Product sub-category slug (see below) |
 | `type_text` | list[str] | Natural language description of each event |
 
-## Event Types (8 categories)
+## Event Types (54 sub-categories)
+
+Event types are the 2nd-level product sub-categories, normalized to lowercase slugs. Rare sub-categories are mapped to `"other"` during initial curation, and all `"other"` events are then stripped from the final sequences. Examples:
 
-| Label | Description |
+| Category | Example Sub-categories |
 |---|---|
-| `issue_opened` | New issue filed |
-| `issue_closed` | Issue resolved/closed |
-| `pr_opened` | Pull request opened |
-| `pr_merged` | Pull request merged |
-| `push` | Code pushed (with commit messages) |
-| `release` | Release published |
-| `pr_reviewed` | PR approved or changes requested |
-| `comment` | Comment on issue, PR, or review |
+| Electronics | `computers_accessories`, `camera_photo`, `television_video`, `headphones_earbuds_accessories`, `home_audio` |
+| Books | `literature_fiction`, `mystery_thriller_suspense`, `children_s_books`, `teen_young_adult`, `romance` |
+| Home & Kitchen | `kitchen_dining`, `home_d_cor_products`, `bedding`, `storage_organization`, `furniture` |
+| Beauty & Personal Care | `hair_care`, `skin_care`, `foot_hand_nail_care`, `makeup`, `tools_accessories` |
 
 ## Curation Filters
 
-Sequences are selected to represent moderately active developers working across a focused set of repositories:
+Sequences are selected to represent moderately prolific reviewers with diverse, non-trivial sub-category distributions:
 
 | Filter | Value | Description |
 |---|---|---|
-| `min-raw-events` | 50 | Min kept-type events in archive (pre-classification) |
-| `max-raw-events` | 300 | Max kept-type events in archive (pre-classification) |
-| `min-events` | 80 | Min classified events per sequence |
-| `max-events` | 100 | Max classified events (skip, not truncate) |
-| `min-unique-types` | 5 | At least 5 distinct event types |
-| `min-repos` | 2 | Active in at least 2 repos |
-| `max-repos` | 5 | Active in at most 5 repos |
-| `min-avg-text-len` | 150 | Min average text length per event |
-| `max-type-fraction` | 0.5 | No single event type >50% of events |
-| `require-english` | true | All event texts must be in English (non-Latin script detection) |
-
-Additional processing:
-- Bot users are excluded (detected by name patterns like `[bot]`, `dependabot`, etc.)
-- Consecutive duplicate events (same type + text) are deduplicated
-
-## Key Difference from Repository Events
-
-| | User Events | Repo Events |
+| `min-events` | 80 | Min reviews per user sequence |
+| `max-events` | 100 | Max reviews per user sequence |
+| `min-types` | 3 | At least 3 distinct sub-category types |
+| `min-span` | 3 months | Exclude sequences spanning less than 3 months |
+| `max-span` | 60 months | Exclude sequences spanning 60+ months |
+| `max-sub-categories` | 15 | Top 15 sub-categories kept per parent category; rest mapped to `"other"` |
+| `min-subcat-count` | 50 | Sub-categories with fewer than 50 products mapped to `"other"` |
+
+### Anti-Exploitation Filters
+
+These filters are applied after initial curation to prevent models from exploiting simple statistical shortcuts:
+
+| Filter | Value | Description |
 |---|---|---|
-| Grouping | By developer | By repository |
-| Cross-repo | Yes (2-5 repos) | No (single repo) |
-| Text perspective | Mentions repository name | Mentions actor name |
-| Use case | Developer workflow modeling | Project activity modeling |
+| Remove "other" events | — | All events with `type_event == "other"` are stripped; `time_since_last_event` is recomputed |
+| Re-check min-events/min-types | 80 / 3 | Sequences re-validated after "other" removal |
+| `max-dominant-ratio` | 0.35 | Drop sequences where any single type exceeds 35% of events |
+| `max-repeat-ratio` | 0.30 | Drop sequences where consecutive same-type rate exceeds 30% |
+
+**Filtering pipeline:** 2,537 → remove "other" events → re-check length (−585) and type diversity (−11) → dominant ratio (−1,550) → repeat ratio (−105) → **286 sequences**.
 
 ## Example
 
 ```json
 {
   "seq_idx": 0,
-  "seq_len": 90,
-  "description": "GitHub user alice: active in org/repo-a (A web framework for Rust), org/repo-b Event window: June 02 - 07, 2024.",
-  "metadata": "{\"actor_login\": \"alice\", \"num_repos\": 2, \"top_repos\": [\"org/repo-a\", \"org/repo-b\"], \"time_unit\": \"hours\"}",
-  "time_since_start": [0.0, 0.004, ...],
-  "time_since_last_event": [0.0, 0.004, ...],
-  "type_event": ["pr_reviewed", "pr_merged", "push", ...],
-  "type_text": ["PR approved in org/repo-a: Add new feature...", ...]
+  "seq_len": 98,
+  "type_category": "Electronics",
+  "span_weeks": 61.33,
+  "description": "Amazon Electronics review timeline for a user spanning Feb 2022 to Apr 2023. This sequence tracks the user's product review activity within the Electronics category on Amazon.",
+  "metadata": "{\"user_id\": \"AHJKYQVJ...\", \"parent_category\": \"Electronics\", \"num_reviews\": 100, \"num_sub_categories\": 12, \"date_range_start\": \"2022-02-17T08:05:52.291000+00:00\", \"date_range_end\": \"2023-04-22T15:21:09.246000+00:00\", \"span_weeks\": 61.33, \"time_unit\": \"weeks\"}",
+  "time_since_start": [0.0, 29.24, 31.50, ...],
+  "time_since_last_event": [0.0, 29.24, 2.26, ...],
+  "type_event": ["camera_photo", "portable_audio_video", "video_projectors", ...],
+  "type_text": ["Product: \"Selfie Ring Light with Tripod Stand...\". 5/5 stars...", ...]
 }
 ```
 
 ## Intended Use
 
 - Training and evaluating temporal point process models
-- Studying individual developer workflow patterns across repos
+- Studying consumer purchasing patterns across product sub-categories
 - Benchmarking next-event prediction and event forecasting
 - Modeling marked temporal point processes with rich text marks
 
@@ -147,11 +113,11 @@ Additional processing:
 If you use this dataset, please cite:
 
 ```bibtex
-@dataset{github_user_events_2024,
-  title={GitHub User Event Streams},
+@dataset{amazon_review_events_2024,
+  title={Amazon Product Review Events},
   author={XiaoBB},
   year={2024},
-  url={https://huggingface.co/datasets/XiaoBB/github_user_events},
-  note={Curated from GH Archive, June 2024}
+  url={https://huggingface.co/datasets/XiaoBB/amazon_review_events},
+  note={Curated from Amazon Reviews 2023 (McAuley Lab)}
 }
 ```