Publications
Graph Edit Distance with General Costs Using Neural Set Divergence
Eeshaan Jain*, Indradyumna Roy*, Saswat Meher, Soumen Chakrabarti, Abir De
NeurIPS 2024
GraphEdx is the first-of-its-kind neural GED framework that incorporates variable edit costs, capable of modeling both symmetric and asymmetric graph (dis)similarities, allowing for more flexible and accurate GED estimation compared to earlier methods.
Iteratively Refined Early Interaction Alignment for Subgraph Matching based Graph Retrieval
Ashwin Ramachandran, Vaibhav Raj, Indradyumna Roy, Soumen Chakrabarti, Abir De
NeurIPS 2024
EinsMatch is an early interaction graph neural network, where the approximate injective alignments between any given graph pair gets progressively refined with successive rounds, resulting in significantly better retrieval performance than existing methods.
Locality Sensitive Hashing in Fourier Frequency Domain For Soft Set Containment Search
Indradyumna Roy, Rishi Agarwal, Soumen Chakrabarti, Anirban Dasgupta, Abir De
NeurIPS 2023 (Spotlight)
FourierHashNet is an asymmetric LSH for hinge distance, which first transforms the hinge distance into a bounded dominance similarity measure, which is then Fourier-transformed into an expectation of inner products of functions in the frequency domain. Finally, the expectations are approximated with an importance-sampled estimate, which allows for the use of traditional Random-Hyperplanes LSH.
Maximum Common Subgraph Guided Graph Retrieval: Late and Early Interaction Networks
Indradyumna Roy, Soumen Chakrabarti, Abir De
NeurIPS 2022
We propose neural architectures for two distinct variants of the MCS metric. The customized late interaction models for each variant, outperform SOTA in terms of retrieval accuracy and speed. Furthermore, an unified early interaction network is proposed, which works well for both variants and affords an additional boost in accuracy at the cost of some retrieval speed.
Interpretable Neural Subgraph Matching for Graph Retrieval
Indradyumna Roy, Venkata Sai Velugoti, Soumen Chakrabarti, Abir De
AAAI 2022
ISONET proposes a novel interpretable neural edge alignment formulation, which enables identification of the underlying subgraph in a corpus graph, which is relevant (isomorphic) to the given query graph. Training for ISONET is done using only binary relevance lavels on graph pairs, without any fine-grained ground truth information about node or edge alignments.
Adversarial Permutation Guided Node Representations for Link Prediction
Indradyumna Roy, Abir De, Soumen Chakrabarti
AAAI 2021
PermGNN casts the link prediction objective as an adversarial game, which allows for usage of order-sensitive RNNs as neighborhood feature aggregators. We ensure permutation insensitivity by optimizing a min-max ranking loss function with respect to the smooth surrogates of adversarial permutations.
Plagiarism Detection In Polyphonic Music Using Monaural Signal Separation
Soham De, Indradyumna Roy, Tarunima Prabhakar, Kriti Suneja, Sourish Chaudhuri, Rita Singh, Bhiksha Raj
INTERSPEECH 2012
We present a novel feature space for audio derived from compositional modelling techniques, commonly used in signal separation, that provides a mechanism to account for polyphony without incurring an inordinate amount of computational overhead.
Eeshaan Jain*, Indradyumna Roy*, Saswat Meher, Soumen Chakrabarti, Abir De
NeurIPS 2024
GraphEdx is the first-of-its-kind neural GED framework that incorporates variable edit costs, capable of modeling both symmetric and asymmetric graph (dis)similarities, allowing for more flexible and accurate GED estimation compared to earlier methods.
Ashwin Ramachandran, Vaibhav Raj, Indradyumna Roy, Soumen Chakrabarti, Abir De
NeurIPS 2024
EinsMatch is an early interaction graph neural network, where the approximate injective alignments between any given graph pair gets progressively refined with successive rounds, resulting in significantly better retrieval performance than existing methods.
Indradyumna Roy, Rishi Agarwal, Soumen Chakrabarti, Anirban Dasgupta, Abir De
NeurIPS 2023 (Spotlight)
FourierHashNet is an asymmetric LSH for hinge distance, which first transforms the hinge distance into a bounded dominance similarity measure, which is then Fourier-transformed into an expectation of inner products of functions in the frequency domain. Finally, the expectations are approximated with an importance-sampled estimate, which allows for the use of traditional Random-Hyperplanes LSH.
Indradyumna Roy, Soumen Chakrabarti, Abir De
NeurIPS 2022
We propose neural architectures for two distinct variants of the MCS metric. The customized late interaction models for each variant, outperform SOTA in terms of retrieval accuracy and speed. Furthermore, an unified early interaction network is proposed, which works well for both variants and affords an additional boost in accuracy at the cost of some retrieval speed.
Indradyumna Roy, Venkata Sai Velugoti, Soumen Chakrabarti, Abir De
AAAI 2022
ISONET proposes a novel interpretable neural edge alignment formulation, which enables identification of the underlying subgraph in a corpus graph, which is relevant (isomorphic) to the given query graph. Training for ISONET is done using only binary relevance lavels on graph pairs, without any fine-grained ground truth information about node or edge alignments.
Indradyumna Roy, Abir De, Soumen Chakrabarti
AAAI 2021
PermGNN casts the link prediction objective as an adversarial game, which allows for usage of order-sensitive RNNs as neighborhood feature aggregators. We ensure permutation insensitivity by optimizing a min-max ranking loss function with respect to the smooth surrogates of adversarial permutations.
Soham De, Indradyumna Roy, Tarunima Prabhakar, Kriti Suneja, Sourish Chaudhuri, Rita Singh, Bhiksha Raj
INTERSPEECH 2012
We present a novel feature space for audio derived from compositional modelling techniques, commonly used in signal separation, that provides a mechanism to account for polyphony without incurring an inordinate amount of computational overhead.