生成式推荐模型的长序列特征：离线存储

长序列特征的例子

For example, a user’s sequence could look like this:
[ Electronics, Clothing, Books, Home & Kitchen, Electronics, Books, Electronics, Sports & Outdoors, Electronics ]
The interactions could be further refined by adding the type of behavior (e.g., [Electronics:view, Clothing:click, Books:purchase, Home & Kitchen:add_to_cart, …] ).

1. Event-level features

Categorical Encoding: Convert event types (e.g., “click”, “add to cart”, “purchase”, “view”) or item categories into numerical representations using techniques like one-hot encoding or embedding methods.
Temporal Features: Extract time-based features from timestamps, such as hour of day, day of the week, month, and time elapsed since previous interaction.
Interaction-Specific Features: Capture attributes specific to each interaction, like product price, rating, duration of video watched, etc.

2. Sequence-level features

Aggregation Features

Count of specific events: Number of clicks, purchases, or searches in the past week.
Average value of numerical features: Average product price of items viewed or purchased.
Time-based statistics: Maximum, minimum, or average time between consecutive interactions.
Frequency of interactions: Number of interactions per hour or day.

Session-based Features

Session length: Number of events or duration of the session.
Session activity type: Percentage of clicks, purchases, or searches within the session.
Sequence of items/events within a session: Representing the order of actions taken by the user, for example, viewing product A, then B, then adding B to the cart.

Temporal Order Features

Lag features: Features from previous interactions (e.g., the last item viewed, the type of the second-to-last event). GeeksforGeeks notes that lag features are a fundamental technique for time-series data.
Positional embeddings: Add positional information to sequence embeddings to capture the order of events.

3. User-level features

Long-Term Preference Features: Summarize user preferences over a long period:
Most frequently purchased categories: Top categories a user interacts with.
Overall spending patterns: Average purchase value, total purchases, etc.
Average interaction count: Average number of interactions per day or week.
User Embeddings

4. Interaction features (between user and item/context)

User-Item Similarity: Calculate the similarity between the current item and previous items the user interacted with.
Time Since Last Interaction with Item: Capturing recency of interest in a particular item.

how to store the long term user behaviro sequence features in offline data lake storage?

1. Schema design： see following
2. File formats：Columnar Formats (Parquet or ORC)
3. Partitioning strategies：Date-Based Partitioning，User ID/Device ID Partitioning
4. Data ingestion and processing：Batch Ingestion，Data Enrichment and Transformation
5. Lifecycle management and cost optimization：Retention Policies

Schema design:

user_id: string
name: string
gender: string
behavior_sequence: array<struct<timestamp: timestamp,category_id: int,action_type: string,product_id: string,price: double>
>

how to update this behavior_sequence field efficiently when there is new behavior for the same user?

Merge Operations (Upserts/MERGE SQL): This allows you to efficiently update existing records (the user_id and its behavior_sequence) and insert new ones

MERGE INTO target_delta_table AS target
USING source_data AS source
ON target.user_id = source.user_id
WHEN MATCHED THENUPDATE SET target.behavior_sequence = array_append(target.behavior_sequence, source.new_behavior)
WHEN NOT MATCHED THENINSERT (user_id, name, behavior_sequence)VALUES (source.user_id, source.name, array(source.new_behavior))

参考资料

和Google的对话记录