Sadamori Kojaku
@skojaku.bsky.social
Assist. Prof. in 🇺🇸 . Interested in Network Science and Computer Science. Love Sailing, Kendo, Tennis, and Cooking Chinese and Italian. My favorite book: 11ぴきのネコ.
GPT Inference (greedy, top-k, beam search, temperature control).
November 29, 2025 at 11:27 AM
GPT Inference (greedy, top-k, beam search, temperature control).
word2vec: from the structural linguistic perspective. Written by me. Illustrated by Nano Banana Pro.
November 27, 2025 at 8:51 PM
word2vec: from the structural linguistic perspective. Written by me. Illustrated by Nano Banana Pro.
Nano Banana Pro unbelievably lowers the entry barrier for people interested in creating short Manga. This blog post describes the recipe (in Japanese). I get this Manga version of my lecture note on the position encoding in transformers, **in the first shot** 😮.
note.com/konho/n/na8d...
note.com/konho/n/na8d...
November 24, 2025 at 6:09 PM
Nano Banana Pro unbelievably lowers the entry barrier for people interested in creating short Manga. This blog post describes the recipe (in Japanese). I get this Manga version of my lecture note on the position encoding in transformers, **in the first shot** 😮.
note.com/konho/n/na8d...
note.com/konho/n/na8d...
Knowledge graph (KG) is often treated as a set of nodes and relations. But relations form motifs that appear frequently in a KG. Learning the motif embedding enhances KG tasks consistently. Need to specify the motifs though the mode is robust to overfitting when too many motifs are given.
July 17, 2025 at 12:19 AM
Knowledge graph (KG) is often treated as a set of nodes and relations. But relations form motifs that appear frequently in a KG. Learning the motif embedding enhances KG tasks consistently. Need to specify the motifs though the mode is robust to overfitting when too many motifs are given.
Matryoshka Embedding is a length adjustable embedding learned by summing the loss computed for sub-vectors of diff lengths truncated from the same embedding. But the quality degeneration is substantial for shorter vectors. CSR alleviates by creating **longer** (> the base emb) but sparse vectors.
July 16, 2025 at 11:12 PM
Matryoshka Embedding is a length adjustable embedding learned by summing the loss computed for sub-vectors of diff lengths truncated from the same embedding. But the quality degeneration is substantial for shorter vectors. CSR alleviates by creating **longer** (> the base emb) but sparse vectors.
@binghamtonu.bsky.social We'll host #ComplexNetworks2025! Please join us on December 9-11 in Binghamton, NY to wrap up 2025 🎅 with researchers on network science, graph machine learning, and complex systems. Submit by Sept 2 for Springer proceedings! 😉 #NetworkScience #BinghamtonUniversity
July 16, 2025 at 4:39 PM
@binghamtonu.bsky.social We'll host #ComplexNetworks2025! Please join us on December 9-11 in Binghamton, NY to wrap up 2025 🎅 with researchers on network science, graph machine learning, and complex systems. Submit by Sept 2 for Springer proceedings! 😉 #NetworkScience #BinghamtonUniversity
Attending #ICML for the first time presenting our work on link prediction! openreview.net/forum?id=K9z...
Reminded my time as a student attending a conference for the first time.
Reminded my time as a student attending a conference for the first time.
July 15, 2025 at 11:39 PM
Attending #ICML for the first time presenting our work on link prediction! openreview.net/forum?id=K9z...
Reminded my time as a student attending a conference for the first time.
Reminded my time as a student attending a conference for the first time.
The disruption index quantifies "bow-tie-ness" in citation networks, i.e., a paper is "disruptive" (controversial interpretation though) if many others cite it instead of its references. Not well-known fact: it is *very expensive to compute*: for a network of 130M nodes, a naive impl. took >50h 😅.
June 10, 2025 at 8:09 PM
The disruption index quantifies "bow-tie-ness" in citation networks, i.e., a paper is "disruptive" (controversial interpretation though) if many others cite it instead of its references. Not well-known fact: it is *very expensive to compute*: for a network of 130M nodes, a naive impl. took >50h 😅.
Quantifying invisible barriers and bridges in city learned from mobility data using neural embedding by @estebanmoro.bsky.social @danielhankim.bsky.social Guanyuan Weng @yyahn.bsky.social #NetSci2025
June 5, 2025 at 2:36 PM
Quantifying invisible barriers and bridges in city learned from mobility data using neural embedding by @estebanmoro.bsky.social @danielhankim.bsky.social Guanyuan Weng @yyahn.bsky.social #NetSci2025
Nicholas Landry introduced an efficient sampling algorithm for microcanonical configuration models of hypergraphs. It avoids loopy edges, generates a new initial sample distinct from empirical data (adhering to realizability conditions), employs a smart criterion for MCMC sampling.
#NetSci2025
#NetSci2025
June 5, 2025 at 10:17 AM
Nicholas Landry introduced an efficient sampling algorithm for microcanonical configuration models of hypergraphs. It avoids loopy edges, generates a new initial sample distinct from empirical data (adhering to realizability conditions), employs a smart criterion for MCMC sampling.
#NetSci2025
#NetSci2025
Ching Jin discusses the recognition process of science, through how scientists climb up the ladder of academic prizes, finding a topological hierarchy of prize mobility, homophily, and provided a minimal model describing the process. #NetSci2025
June 4, 2025 at 3:30 PM
Ching Jin discusses the recognition process of science, through how scientists climb up the ladder of academic prizes, finding a topological hierarchy of prize mobility, homophily, and provided a minimal model describing the process. #NetSci2025
Which collaboration ties are similarity-driven or complementarity-driven? Elizaveta Evmenova talked about how to identify them, motifs of each type, fragmentations, showing that similarity ties are local while complementary are long range. #NetSci2025
June 4, 2025 at 2:58 PM
Which collaboration ties are similarity-driven or complementarity-driven? Elizaveta Evmenova talked about how to identify them, motifs of each type, fragmentations, showing that similarity ties are local while complementary are long range. #NetSci2025
Diego Garlaschelli discusses when the micro canonical and canonical models are asymptomatically equivalent, and which one is favored in terms of description length. In a nutshell, they are equivalent when constrains are global (eg the # of edges) but not for local (eg deg. seq. ) #NetSci2025
June 4, 2025 at 1:43 PM
Diego Garlaschelli discusses when the micro canonical and canonical models are asymptomatically equivalent, and which one is favored in terms of description length. In a nutshell, they are equivalent when constrains are global (eg the # of edges) but not for local (eg deg. seq. ) #NetSci2025
The optimization is often under-appreciated but important factor underpinning the performance of community detection. Samin showcased in case of the modularity maximization (but not limited to)
June 4, 2025 at 1:18 PM
The optimization is often under-appreciated but important factor underpinning the performance of community detection. Samin showcased in case of the modularity maximization (but not limited to)
Marcella Tambuscio talked about digitizing history, focusing the time of Maximillian, the first emperor of Holy Roman Empire 500y ago!). Interesting networks, eg churches and families, approaching quantitatively to how and why Maximillian established the prosperity of Habsburg family #NetSci2025
June 4, 2025 at 10:20 AM
Marcella Tambuscio talked about digitizing history, focusing the time of Maximillian, the first emperor of Holy Roman Empire 500y ago!). Interesting networks, eg churches and families, approaching quantitatively to how and why Maximillian established the prosperity of Habsburg family #NetSci2025
@complenet.bsky.social Now that the conference is over. It,s been inspirational, with many inspiring talks. Cachaça party made it more special 🥃. Thanks to the organizers for the smooth operation and for overcoming many challenges, from distance to political hurdles 😼
April 26, 2025 at 12:17 AM
@complenet.bsky.social Now that the conference is over. It,s been inspirational, with many inspiring talks. Cachaça party made it more special 🥃. Thanks to the organizers for the smooth operation and for overcoming many challenges, from distance to political hurdles 😼
@complenet.bsky.social
Inspirational talk by Esteban Moro on urban mobility, with fun trivia, causal insights, and the idea of “mobility genes.” Yes. I spend most of my time in just five places🫣, yet it’s the rare exceptions, like this conference, that stands out most in memory.
Inspirational talk by Esteban Moro on urban mobility, with fun trivia, causal insights, and the idea of “mobility genes.” Yes. I spend most of my time in just five places🫣, yet it’s the rare exceptions, like this conference, that stands out most in memory.
April 23, 2025 at 9:31 PM
@complenet.bsky.social
Inspirational talk by Esteban Moro on urban mobility, with fun trivia, causal insights, and the idea of “mobility genes.” Yes. I spend most of my time in just five places🫣, yet it’s the rare exceptions, like this conference, that stands out most in memory.
Inspirational talk by Esteban Moro on urban mobility, with fun trivia, causal insights, and the idea of “mobility genes.” Yes. I spend most of my time in just five places🫣, yet it’s the rare exceptions, like this conference, that stands out most in memory.
I grew up drinking a local Japanese Guarana, sold only in a specific prefecture. Years later in Brazil, I was surprised to encounter Brazilian Guarana—different, but just as delicious.
April 21, 2025 at 6:58 PM
I grew up drinking a local Japanese Guarana, sold only in a specific prefecture. Years later in Brazil, I was surprised to encounter Brazilian Guarana—different, but just as delicious.
My town is nearly empty now. Feel left behind, including almost every playground equipment for us to play😎
June 18, 2024 at 8:23 PM
My town is nearly empty now. Feel left behind, including almost every playground equipment for us to play😎
Found this chain (network) of gears. The three gears have handles and rotating one turns all gears, but different forces. The one at the center needs least force and the one at the right most needs the most. Why? 🤔💭
June 7, 2024 at 11:17 PM
Found this chain (network) of gears. The three gears have handles and rotating one turns all gears, but different forces. The one at the center needs least force and the one at the right most needs the most. Why? 🤔💭
My another office today.
June 3, 2024 at 3:24 PM
My another office today.