package graph_mining

Get desktop application:
View/edit binary Protocol Buffers messages

Configuration proto for ConvertToUndirected. Next available id: 5

• optional ConvertToUndirectedConfig.AsymmetricEdgeTreatment asymmetric_edge_treatment = 1

• oneof edge_set_transformation

  Specifies how to transform the edge set into an undirected graph. Setting no option is equivalent to setting remove_one_sided_edges to false.

  • bool remove_one_sided_edges = 2

    If set to false, the graph is made undirected by adding reciprocal edges. Otherwise, edges with no reciprocal edges are removed. In *both* cases the output contains a Node for each endpoint of each edge in the input. In particular, we may add a Node that ends up being isolated in the output.

  • ConvertToUndirectedConfig.Sparsify sparsify = 3

    See the comments for Sparsify message above.

• optional string counter_prefix = 4

  Prefix for Flume counters.

Specifies how to compute the edge weight of an edge x -> y, given the weights of edges x -> y and y -> x. Most of the options are self-descriptive.

Used in: ConvertToUndirectedConfig

• MAX = 0

• MIN = 1

• REMOVE = 2

  If the weights differ by more than 1e-6, remove the edges. Otherwise use one of the two weights.

• AVG = 3

• SUM = 4

Used in: ConvertToUndirectedConfig

• optional int32 soft_degree_threshold = 1

  Let s = soft_degree_threshold and h = hard_degree_threshold. The following steps are performed. 1. If s is set, for each node we discard all but s edges with highest weight (using node id hashes for tiebreaking). 2. We convert the graph to an undirected one, by adding reciprocal edges. 3. If h is set, we repeat (1) using h as a threshold and then make the graph undirected by removing one sided edges. At least one of *_degree_threshold has to be set.

• optional int32 hard_degree_threshold = 2

• optional bool keep_lowest_weight_edges = 3

  By default the highest weight edges are keep unless keep_lowest_weight_edges is set to true.

Config to employ dynamically changing weight threshold in each round of affinity clustering. Next available id: 4

Used in: in_memory.AffinityClustererConfig

• optional DynamicWeightThresholdConfig.WeightDecayFunction weight_decay_function = 1

  Specifies how the weight threshold changes each round. NOTE: originally this only supported decreasing edge weights in each step (hence the 'decay' in the name, and the threshold in the first step is 'upper_bound'). However, *increasing* weights in each round is also supported and can be achieved by setting lower_bound > upper_bound.

• optional double upper_bound = 2

  Upper bound is the weight threshold that will be used in the first round.

• optional double lower_bound = 3

  Lower bound is the weight threshold that will be used in the last round.

Used in: DynamicWeightThresholdConfig

• UNKNOWN_FUNCTION = 0

• LINEAR_DECAY = 1

  Changes the weight threshold linearly each iteration.

• EXPONENTIAL_DECAY = 2

  Changes the weight threshold exponentially each iteration. Each iteration, a new weight threshold is computed by the following formula: upper_bound*((lower_bound/upper_bound)^(round/compression_rounds))

Config message for GenerateFixedDimensionalEncoding function.

• optional int32 dimension = 3

  Dimension of the input embeddings.

• optional int32 num_repetitions = 4

  Number of independent repetitions for FDE generation.

• optional int32 num_simhash_projections = 5

  Number of SimHash projections used to partition space in each repetition. Is ignored if a partitioning method that is not "DEFAULT_SIMHASH" is being used.

• optional int32 seed = 6

  Seed for the FDE generation process. Must be set to the same value for query and document FDE generation to ensure coonsistency of the partitions and projections used for queries and docuemnets.

• optional FixedDimensionalEncodingConfig.EncodingType encoding_type = 7

• optional int32 projection_dimension = 8

  If a random projection `encoding_type` is being used, this is the dimension to which points are reduced via random projections.

• optional FixedDimensionalEncodingConfig.ProjectionType projection_type = 9

• optional bool fill_empty_partitions = 10

  If true, the blocks of the output FDE corresponding to an empty partition are filled with the coordinates of the point in the input point cloud which is nearest to the partition. Here, the nearest point is defined as the point which minimizes the number of disagreeing bits between its SimHash sketch and the bits in the sketch corresponding to the partition. This option throws an InvalidArgumentError if enabled when used with query FDE generation.

• optional int32 final_projection_dimension = 11

  If set, the final FDE is projected down to this dimension using a random projection. The random projection is implemented using Count-Sketch (i.e., sparse Johnson-Lindenstrauss projection) for efficiency purposes, allowing the original FDE to be constructed with large dimension before being projected down.

How embeddings are added to the FDE. `DEFAULT_SUM` means that points are summed into the respective locations in the FDE. `AVERAGE` means that an average of all embeddings mapped to a partition in the FDE is taken. Generally, Query FDE Encoodings will use `DEFAULT_SUM`, whereas document side FDE generation will use `AVERAGE`.

Used in: FixedDimensionalEncodingConfig

• DEFAULT_SUM = 0

• AVERAGE = 1

The ProjectionType sets how the original embeddings are projected down to `projection_dimension` dimensional space before being added to the FDE. -- `DEFAULT_IDENTITY` means that no projection is set, and the original embeddings are added to the FDE. -- `AMS_SKETCH` a dense AMS sketch is used to project the data original embeddings.

Used in: FixedDimensionalEncodingConfig

• DEFAULT_IDENTITY = 0

• AMS_SKETCH = 1