Catalog

[[TOC]]

SpnWrapper

This is a database specific wrapper around sum-product networks (SPNs). An SpnWrapper can be seen as a relational sum-product network (RSPN) but the terms SPN and RSPN are often used interchangeably. The documentation of SPNs can be found here. Most methods are the same with the difference that some access the attributes (random variables) via attribute name of the table instead of attribute id. This is done via the AttrFilter typename. The AttrFilter typename is defined as follows:

std::unordered_map<ThreadSafePooledString, std::pair<Spn::SpnOperator, float>>

The difference to Filter is that we map from a ThreadSafePooledString representing the attribute name.

The following method can be used to learn an RSPN directly on a table:

static SpnWrapper learn_spn_table(
    const ThreadSafePooledString &name_of_database,
    const ThreadSafePooledString &name_of_table,
    std::vector<Spn::LeafType> leaf_types = decltype(leaf_types)()
);

You can specify the leaf_types or leave them out for automatic detection (DISCRETE for integer types, CONTINUOUS for floats).

The following method can be used to learn an RSPN on each table in a database:

static std::unordered_map<ThreadSafePooledString, SpnWrapper*> learn_spn_database(
    const ThreadSafePooledString &name_of_database,
    std::unordered_map<ThreadSafePooledString, std::vector<Spn::LeafType>> leaf_types = decltype(leaf_types)()
);

You can specify the leaf_types for each table in a map from table name to leaf_types vector. If a table is not in the map, the method uses automatic detection (leave map out for automatic detection for all tables). It returns a map from table name to the respective RSPN.