### Table 1: Symbols associated with the author topic model, as used in this paper. Authors of the corpus A Set Authors of the dth document

2005

"... In PAGE 7: ... In the results in this paper we assume that these hyperparameters are xed. Table1 summarizes this notation. The sequential procedure of rst picking an author followed by picking a topic then generating a word according to the probability distributions above leads to the following generative process:... In PAGE 11: ... Similarly CWT is the word-topic count matrix, where CWT wt is the number of words from the wth entry in the vocabulary assigned to topic t. The other symbols are summarized in Table1 . This equation can be manipulated further to obtain the conditional probability of the topic of the ith word given the rest, P (zi = tjz i; x; Dtrain ; ), and for the conditional probability of the author of the ith word given the rest, P (xi = ajz; x i; Dtrain ; ).... ..."

### Table 2 shows topic and author distributions for three communities trained with different data: one on authors, one on topics, and one on both. The first two methods cor- respond to eliminating either the second or third terms of the

### Table 4: Author-Persona-Topic distributions for David Karger, sorted by the number of papers per persona. For comparison, the categories Karger lists on his publications web page include \Cuts and Flows, quot; \Appli- cations of Theory quot; (which includes the Chord peer-to-peer lookup protocol), \Information Retrieval, quot; and \Graph Coloring. quot; The number on the left is the number of words assigned to each topic within the persona.

"... In PAGE 6: ... The personas discovered by the APT model are also co- herent and interpretable. Examples of personas for two Computer Science researchers are shown in Table4 (David Karger) and Table 5 (Daphne Koller). We also list in the captions of those tables subject terms that the researchers themselves chose for their own papers, as listed on their web pages.... ..."

### Table 5: Author-Persona-Topic distributions for Daphne Koller, sorted by the number of papers per persona. Koller annotates papers on her publications web page with topical labels. These include \Bayesian Networks, quot; \Computational Game Theory, quot;\Computational Biology, quot;\Learning Graphical Models, quot;\Natural Language, quot; \Text and Web quot; and \Theoretical Computer Science quot;

"... In PAGE 6: ... The personas discovered by the APT model are also co- herent and interpretable. Examples of personas for two Computer Science researchers are shown in Table 4 (David Karger) and Table5 (Daphne Koller). We also list in the captions of those tables subject terms that the researchers themselves chose for their own papers, as listed on their web pages.... ..."

### Table 4: Author-Persona-Topic distributions for David Karger, sorted by the number of papers per persona. For comparison, the categories Karger lists on his publications web page include \Cuts and Flows, quot; \Appli- cations of Theory quot; (which includes the Chord peer-to-peer lookup protocol), \Information Retrieval, quot; and \Graph Coloring. quot; The number on the left is the number of words assigned to each topic within the persona.

"... In PAGE 6: ... The personas discovered by the APT model are also co- herent and interpretable. Examples of personas for two Computer Science researchers are shown in Table4 (David Karger) and Table 5 (Daphne Koller). We also list in the captions of those tables subject terms that the researchers themselves chose for their own papers, as listed on their web pages.... ..."

### Table 5: Author-Persona-Topic distributions for Daphne Koller, sorted by the number of papers per persona. Koller annotates papers on her publications web page with topical labels. These include \Bayesian Networks, quot; \Computational Game Theory, quot;\Computational Biology, quot;\Learning Graphical Models, quot;\Natural Language, quot; \Text and Web quot; and \Theoretical Computer Science quot;

"... In PAGE 6: ... The personas discovered by the APT model are also co- herent and interpretable. Examples of personas for two Computer Science researchers are shown in Table 4 (David Karger) and Table5 (Daphne Koller). We also list in the captions of those tables subject terms that the researchers themselves chose for their own papers, as listed on their web pages.... ..."

### Table 6: The best mean reciprocal rank (MRR) score for GCA with discriminative learning, GCA with regular learning, the author-topic (AT) model and SVD. We also include the MRR score of GCA on the word part of the data set only to show the benefit of including additional information from non-word modalities. All models are trained with 20 hidden topics.

"... In PAGE 8: ... We also compare our model versus SVD: we ignore the heterogeneity of the data, and aggregate the word counts, the authors, and the timestamps of the documents into a big matrix, conduct SVD analysis, and then find the lower dimensional representations of the documents. The MRR scores are shown in Table6 . To demonstrate the advantage of incorporating information from multiple modalities, we also run the model on words only.... In PAGE 8: ... To demonstrate the advantage of incorporating information from multiple modalities, we also run the model on words only. As ob- served in Table6 , even with regular training, our model outperforms SVD and AT with uniform prior on authors (not shown in Table 6). With discriminative training, our model is significantly better than SVD and the author-topic model, achieving a MRR score more than twice as large as from SVD.... In PAGE 8: ... To demonstrate the advantage of incorporating information from multiple modalities, we also run the model on words only. As ob- served in Table 6, even with regular training, our model outperforms SVD and AT with uniform prior on authors (not shown in Table6 ). With discriminative training, our model is significantly better than SVD and the author-topic model, achieving a MRR score more than twice as large as from SVD.... In PAGE 9: ... We randomly split the Email data set into training set (9/10, 4,179 documents) and test set (1/10, 464 documents). The MRR scores are reported in Table6 . Again, we can quickly see that the discriminatively trained GCA greatly outperforms other models.... ..."

### Table 1: Topic distributions for the community with the largest number of papers by Jordan M in three models, one trained only on authors, one only on topics, and one on both. The author-based model clusters Jordan and his coauthors, while the topic-based models distinguish between different areas of research.

### Table 2. Top authors in the topic data manage- ment when m = 2, n = 2, k = 1. con# is the number of neighbors in the social network; p# is the number of papers; cite# is the num- ber of citations; r denotes the ranks by the corresponding methods.

2007

"... In PAGE 8: ...verage improvement of 27.8%, 19.1%, 10.6%, and 7.7% over rankings by the number of papers, the topic weights, the number of citations, and the PageRank. We list the top 15 authors ordered by the Co-Ranking scores on the topics data management and learning and classifications in Table2 and Table 3. Along with both ta- bles, the ranks based on simple metrics are also presented.... ..."

Cited by 1