CIKM17: temporally like-minded user community identification through neural embeddings
We propose a neural embedding approach to identify temporally like-minded user communities, i.e., those communities of users who have similar temporal alignment in their topics of interest. Like-minded user communities in social networks are usually identified by either considering explicit structural connections between users (link analysis), users topics of interest expressed in their posted contents (content analysis), or in tandem. In such communities, however, the users rich temporal behavior towards topics of interest is overlooked. Only few recent research efforts consider the time dimension and define like-minded user communities as groups of users who share not only similar topical interests but also similar temporal behavior. Temporal like-minded user communities find application in areas such as recommender systems where relevant items are recommended to the users at the right time. In this paper, we tackle the problem of identifying temporally like-minded user communities by leveraging unsupervised feature learning (embeddings). Specifically, we learn a mapping from the user space to a low-dimensional vector space of features that incorporate both topics of interest and their temporal nature. We demonstrate the efficacy of our proposed approach on a Twitter dataset in the context of three applications: news recommendation, user prediction and community selection, where our work is able to outperform the state-of-the-art on important information retrieval metrics.
![8
regions of like-mindedness (RoL)
Zhao et al. (TriCluster) in genes 3d microarray
equality is [0,
0.1)
equality is [0.1, 1.0]](https://image.slidesharecdn.com/cikm17-171231181828/85/CIKM17-temporally-like-minded-user-community-identification-through-neural-embeddings-8-320.jpg)
![9
r = {u1,u2,u3} {}, C =[z40, z40, z41, z41, ..., z45, z45
r = {u1,u2,u3}{{} + z40}, C = [z40, z41, z41, ..., z45,
z45]
r= {u1,u2}{z40+z40}, C =[z41, z41, ..., z45, z45]](https://image.slidesharecdn.com/cikm17-171231181828/85/CIKM17-temporally-like-minded-user-community-identification-through-neural-embeddings-9-320.jpg)
CIKM17: temporally like-minded user community identification through neural embeddings
- 1. 1
- 2. different temporal inclination of three Twitter users with respect to the War in Afghanistan 2
- 3. 3 motivation: War in Afghanistan item recommendation with correct timing hypothesis: like-minded users exhibit similar temporal behavior towards similar topics due to sth
- 4. 4 Hu et al. aaai'14 Group Specific Topics-over-Time (GrosToT) thank you so much for your clean implementation much appreciated! Fani et al. ci15 user-topic timeseries 2d-xcorrelation graph clustering
- 5. 5 Gold Standard approach 1. regions of like-mindedness (RoL) 2. embeddings 3. graph clustering
- 6. 6 Gold Standard approach 1. regions of like-mindedness (RoL) identify the co-occurrence context of users in topic and time spaces user-topic-time cuboids
- 7. 7 Gold Standard approach 1. regions of like-mindedness (RoL) 1. for each time: 2d RoLs in user and topic spaces 1. build a multigraph Gt = (V, E) V = topics E = {Utzi ,zj(c): to be the maximal set of users whose interest towards zi and zj satisfies the condition of homogeneity c.} 2. dfs 2. for each 2d RoLs: 3d RoLs in (user,topic) and time spaces
- 8. 8 regions of like-mindedness (RoL) Zhao et al. (TriCluster) in genes 3d microarray equality is [0, 0.1) equality is [0.1, 1.0]
- 9. 9 r = {u1,u2,u3} {}, C =[z40, z40, z41, z41, ..., z45, z45 r = {u1,u2,u3}{{} + z40}, C = [z40, z41, z41, ..., z45, z45] r= {u1,u2}{z40+z40}, C =[z41, z41, ..., z45, z45]
- 10. 10 Gold Standard approach 1. regions of like-mindedness (RoL) identify the co-occurrence context of users in topic and time spaces user-topic-time cuboids 2. embeddings input the user space of the RoL to w2v (cbow) and build u2v 3. graph clustering
- 11. 11
- 12. 12 Gold Standard approach 1. regions of like-mindedness (RoL) identify the co-occurrence context of users in topic and time spaces user-topic-time cuboids 2. embeddings input PoTI to w2v and build u2v 3. graph clustering Louvain method on weighted graph based on u2v cosine similarity
- 13. 13 Gold Standard gold standard assumption: users are interested in the topics of the news article about which they have posted golden set: news articles to which a user has explicitly linked in her tweets mentions = {(user, news article, timestamp)} Abel et al.: Twitter, 3M tweets posted by 135K users between Nov. 1 and Dec. 31, 2010. 25,756 triples extracted from 3,468 distinct news articles posted by 1,922 users
- 14. 14 Gold Standard evaluation 1. news recommendation: at time t, recommend news article a to all communities recommendation task: {(user, ?, timestamp) prediction task: (?, news article, timestamp) 2. community selection given a news article a at time t (the input query), find the communities of those users (similar to documents related to an input query) who have mentioned the news article at that time
- 16. 16 temporallylike-minded user community identification through neural embeddings laboratory for systems, software and semantics (LS3) @UNB @RyersonU
Editor's Notes
- #10: To find the final 2-d RoLs for time t , we apply depth-first-search (DFS) on the multigraph Gt based on the pseudo code described in Algorithm 1. We start with a 2-d RoL r = U .; all users U, but no topics since no node (topic) has been processed yet and C = [z1, z1, z2, z2, ..., z |Z| , z |Z| ] as the set of all initial nodes (topics) to be processed. Here,C includes duplicated initial topics to support for directed loops on each node. At each intermediate recursive call, we have a current candidate 2-d RoL r = A B and a list of not yet processed topics C. We add r into an initially empty set Rt if it satisfies c and is not already contained in some RoL r Rt . Then, we remove any 2-d RoL r Rt , which has already been subsumed by r (lines 2-6). We expand the current candidate r from each of its old topics zi to a new topic zj if there is a directed edge (zi zj ) Ut . Then, the function is called on the new candidate {r .A Ut zi ,zj } {r .B {zj }} (lines 7-15). For example, let us consider how the 2-d RoLs are identified from the multigraph G22 shown in Figure 4a. Initially the algorithm starts with the candidate 2-d RoL r = {u1,u2,u3} .,C = [z40, z40, z41, z41, ..., z45, z45]. We pop node z40 and recursively call the function onr = {u1,u2,u3}{z40},C = [z40, z41, z41, ..., z45, z45] (line 10). Since {u1,u2,u3} {z40} does not satisfy condition c, we continue by popping a new node (topic) which is again z40. There is only one directed edge (loop) from z40 z40, so we obtain a new candidate (line 14) and call the function on r = {u1,u2}{z40},C = [z41, z41, ..., z45, z45] (line 15). Now, the input r satisfies c and we add it to the thus far empty R22 (line 6). Next, we pop z41 and there is a directed edge from z40 z41 with U22 z40,z41 = {u2}. So we call the function on r = {u2} {z40, z41},C = [z41, ..., z45, z45] which leads to a new element in R22.