rel-hm: YAML manifests + descriptions + schema card

rel-hm/README.md

@@ -0,0 +1,50 @@
+---
+tags:
+- relbench
+- relational-deep-learning
+pretty_name: rel-hm
+---
+
+# rel-hm
+
+H&M e-commerce: customers, articles, and time-stamped purchase transactions.
+
+## Schema
+
+```mermaid
+erDiagram
+    transactions {
+        key customer_id FK
+        key article_id FK
+        datetime t_dat
+    }
+    article {
+        key article_id PK
+    }
+    customer {
+        key customer_id PK
+    }
+    transactions }o--|| customer : customer_id
+    transactions }o--|| article : article_id
+```
+
+Splits: validation `2020-09-07`, test `2020-09-14` (rows up to a split's timestamp are the inputs for that split).
+
+## Tasks
+
+| task | kind | type | description |
+|---|---|---|---|
+| `item-sales` | forecast | regression | Predict the total sales for an article (the sum of prices of the associated transactions) in the next week. |
+| `transactions-price` | autocomplete | regression | Predict the `price` column of the `transactions` table. |
+| `user-churn` | forecast | binary_classification | Predict the churn for a customer (no transactions) in the next week. |
+| `user-item-purchase` | forecast | link_prediction | Predict the list of articles each customer will purchase in the next seven days. |
+
+## Loading
+
+```python
+import relbench
+ds = relbench.load_dataset("rel-hm")
+task = relbench.load_task("rel-hm", "<task>")
+```
+
+Manifest layout (`manifest.yaml` + plain parquet); see the RelBench [CONTRIBUTING guide](https://github.com/snap-stanford/relbench/blob/main/CONTRIBUTING.md).

rel-hm/manifest.json

@@ -1,26 +0,0 @@
-{
-  "name": "rel-hm",
-  "manifest_version": 1,
-  "val_timestamp": "2020-09-07",
-  "test_timestamp": "2020-09-14",
-  "tables": {
-    "transactions": {
-      "pkey": null,
-      "time_col": "t_dat",
-      "fkeys": {
-        "customer_id": "customer",
-        "article_id": "article"
-      }
-    },
-    "article": {
-      "pkey": "article_id",
-      "time_col": null,
-      "fkeys": {}
-    },
-    "customer": {
-      "pkey": "customer_id",
-      "time_col": null,
-      "fkeys": {}
-    }
-  }
-}

rel-hm/manifest.yaml

@@ -0,0 +1,20 @@
+name: rel-hm
+manifest_version: 1
+description: 'H&M e-commerce: customers, articles, and time-stamped purchase transactions.'
+val_timestamp: '2020-09-07'
+test_timestamp: '2020-09-14'
+tables:
+  transactions:
+    pkey: null
+    time_col: t_dat
+    fkeys:
+      customer_id: customer
+      article_id: article
+  article:
+    pkey: article_id
+    time_col: null
+    fkeys: {}
+  customer:
+    pkey: customer_id
+    time_col: null
+    fkeys: {}

rel-hm/tasks/item-sales/manifest.json

@@ -1,17 +0,0 @@
-{
-  "name": "item-sales",
-  "kind": "forecast",
-  "task_type": "regression",
-  "entity_table": "article",
-  "entity_col": "article_id",
-  "target_col": "sales",
-  "time_col": "timestamp",
-  "timedelta": "7 days",
-  "metrics": [
-    "r2",
-    "mae",
-    "rmse"
-  ],
-  "sql": "\n            SELECT\n                timestamp,\n                article_id,\n                sales\n            FROM\n                timestamps,\n                article,\n                (\n                    SELECT\n                        COALESCE(SUM(price), 0) as sales\n                    FROM\n                        transactions,\n                    WHERE\n                        transactions.article_id = article.article_id AND\n                        t_dat > timestamp AND\n                        t_dat <= timestamp + INTERVAL '{timedelta}'\n                )\n            ",
-  "manifest_version": 1
-}

rel-hm/tasks/item-sales/manifest.yaml

@@ -0,0 +1,11 @@
+name: item-sales
+kind: forecast
+task_type: regression
+description: Predict the total sales for an article (the sum of prices of the associated transactions) in the next week.
+entity_table: article
+entity_col: article_id
+target_col: sales
+time_col: timestamp
+timedelta: 7 days
+sql: "\n            SELECT\n                timestamp,\n                article_id,\n                sales\n            FROM\n                timestamps,\n                article,\n                (\n                    SELECT\n                        COALESCE(SUM(price), 0) as sales\n                    FROM\n                        transactions,\n                    WHERE\n                        transactions.article_id = article.article_id AND\n                        t_dat > timestamp AND\n                        t_dat <= timestamp + INTERVAL '{timedelta}'\n                )\n            "
+manifest_version: 1

rel-hm/tasks/transactions-price/manifest.json

@@ -1,9 +0,0 @@
-{
-  "name": "transactions-price",
-  "kind": "autocomplete",
-  "task_type": "regression",
-  "entity_table": "transactions",
-  "target_col": "price",
-  "metrics": [],
-  "manifest_version": 1
-}

rel-hm/tasks/transactions-price/manifest.yaml

@@ -0,0 +1,7 @@
+name: transactions-price
+kind: autocomplete
+task_type: regression
+description: Predict the `price` column of the `transactions` table.
+entity_table: transactions
+target_col: price
+manifest_version: 1

rel-hm/tasks/user-churn/manifest.json

@@ -1,18 +0,0 @@
-{
-  "name": "user-churn",
-  "kind": "forecast",
-  "task_type": "binary_classification",
-  "entity_table": "customer",
-  "entity_col": "customer_id",
-  "target_col": "churn",
-  "time_col": "timestamp",
-  "timedelta": "7 days",
-  "metrics": [
-    "average_precision",
-    "accuracy",
-    "f1",
-    "roc_auc"
-  ],
-  "sql": "\n            SELECT\n                timestamp,\n                customer_id,\n                CAST(\n                    NOT EXISTS (\n                        SELECT 1\n                        FROM transactions\n                        WHERE\n                            transactions.customer_id = customer.customer_id AND\n                            t_dat > timestamp AND\n                            t_dat <= timestamp + INTERVAL '{timedelta}'\n                    ) AS INTEGER\n                ) AS churn\n            FROM\n                timestamps,\n                customer,\n            WHERE\n                EXISTS (\n                    SELECT 1\n                    FROM transactions\n                    WHERE\n                        transactions.customer_id = customer.customer_id AND\n                        t_dat > timestamp - INTERVAL '{timedelta}' AND\n                        t_dat <= timestamp\n                )\n            ",
-  "manifest_version": 1
-}

rel-hm/tasks/user-churn/manifest.yaml

@@ -0,0 +1,11 @@
+name: user-churn
+kind: forecast
+task_type: binary_classification
+description: Predict the churn for a customer (no transactions) in the next week.
+entity_table: customer
+entity_col: customer_id
+target_col: churn
+time_col: timestamp
+timedelta: 7 days
+sql: "\n            SELECT\n                timestamp,\n                customer_id,\n                CAST(\n                    NOT EXISTS (\n                        SELECT 1\n                        FROM transactions\n                        WHERE\n                            transactions.customer_id = customer.customer_id AND\n                            t_dat > timestamp AND\n                            t_dat <= timestamp + INTERVAL '{timedelta}'\n                    ) AS INTEGER\n                ) AS churn\n            FROM\n                timestamps,\n                customer,\n            WHERE\n                EXISTS (\n                    SELECT 1\n                    FROM transactions\n                    WHERE\n                        transactions.customer_id = customer.customer_id AND\n                        t_dat > timestamp - INTERVAL '{timedelta}' AND\n                        t_dat <= timestamp\n                )\n            "
+manifest_version: 1

rel-hm/tasks/user-item-purchase/manifest.json

@@ -1,20 +0,0 @@
-{
-  "name": "user-item-purchase",
-  "kind": "forecast",
-  "task_type": "link_prediction",
-  "target_col": "article_id",
-  "time_col": "timestamp",
-  "src_entity_table": "customer",
-  "src_entity_col": "customer_id",
-  "dst_entity_table": "article",
-  "dst_entity_col": "article_id",
-  "eval_k": 12,
-  "timedelta": "7 days",
-  "metrics": [
-    "link_prediction_precision",
-    "link_prediction_recall",
-    "link_prediction_map"
-  ],
-  "sql": "\n            SELECT\n                t.timestamp,\n                transactions.customer_id,\n                LIST(DISTINCT transactions.article_id) AS article_id\n            FROM\n                timestamps t\n            LEFT JOIN\n                transactions\n            ON\n                transactions.t_dat > t.timestamp AND\n                transactions.t_dat <= t.timestamp + INTERVAL '{timedelta}'\n            GROUP BY\n                t.timestamp,\n                transactions.customer_id\n            ",
-  "manifest_version": 1
-}

rel-hm/tasks/user-item-purchase/manifest.yaml

@@ -0,0 +1,14 @@
+name: user-item-purchase
+kind: forecast
+task_type: link_prediction
+description: Predict the list of articles each customer will purchase in the next seven days.
+target_col: article_id
+time_col: timestamp
+src_entity_table: customer
+src_entity_col: customer_id
+dst_entity_table: article
+dst_entity_col: article_id
+eval_k: 12
+timedelta: 7 days
+sql: "\n            SELECT\n                t.timestamp,\n                transactions.customer_id,\n                LIST(DISTINCT transactions.article_id) AS article_id\n            FROM\n                timestamps t\n            LEFT JOIN\n                transactions\n            ON\n                transactions.t_dat > t.timestamp AND\n                transactions.t_dat <= t.timestamp + INTERVAL '{timedelta}'\n            GROUP BY\n                t.timestamp,\n                transactions.customer_id\n            "
+manifest_version: 1