Work supported by NSF Grants IIS-0331707 and IIS-0083489 Copyright(c) by Dmitri V. Kalashnikov, 2005 Exploiting Relationships for Object Consolidation Zhaoqi Chen Dmitri V. Kalashnikov Sharad Mehrotra Computer Science Department University of California, Irvine http://www.ics.uci.edu/~dvk/RelDC http://www.itr-rescue.org (RESCUE) ACM IQIS 2005 Talk Overview • Examples – motivating data cleaning (DC) – motivating analysis of relationships for DC • Object consolidation – one of the DC problems – this work addresses • Proposed approach – RelDC framework – Relationship analysis and graph partitioning • Experiments 2 Why do we need “Data Cleaning”? q ??? Hi, my name is Jane Smith. I’d like to apply for a faculty position at your university Publications: 1. …… 2. …… 3. …… Jane Smith – Fresh Ph.D. Wow! Unbelievable! You must be a really hard worker! I amOK, let me sure we will accept check a candidate like something that! quickly … CiteSeer Rank Tom - Recruiter 3 What is the problem? Suspicious entries – Lets go to DBLP website – which stores bibliographic entries of many CS authors – Lets check two people – “A. Gupta” – “L. Zhang” CiteSeer: the top-k most cited authors DBLP DBLP 4 Comparing raw and cleaned CiteSeer Rank Author Location # citations 1 (100.00%) douglas schmidt cs@wustl 5608 2 (100.00%) rakesh agrawal almaden@ibm 4209 3 (100.00%) hector garciamolina @ 4167 4 (100.00%) sally floyd @aciri 3902 5 (100.00%) jennifer widom @stanford 3835 6 (100.00%) david culler cs@berkeley 3619 6 (100.00%) thomas henzinger eecs@berkeley 3752 7 (100.00%) rajeev motwani @stanford 3570 8 (100.00%) willy zwaenepoel cs@rice 3624 9 (100.00%) van jacobson lbl@gov 3468 10 (100.00%) rajeev alur cis@upenn 3577 11 (100.00%) john ousterhout @pacbell 3290 12 (100.00%) joseph halpern cs@cornell 3364 13 (100.00%) andrew kahng @ucsd 3288 14 (100.00%) peter stadler tbi@univie 3187 15 (100.00%) serge abiteboul @inria 3060 Cleaned CiteSeer top-k CiteSeer top-k 5 What is the lesson? “Garbage in, garbage out” principle: Making decisions based on bad data, can lead to wrong results. – – – – – data should be cleaned first e.g., determine the (unique) real authors of publications solving such challenges is not always “easy” that explains a large body of work on data cleaning note – CiteSeer is aware of the problem with its ranking – there are more issues with CiteSeer – many not related to data cleaning 6 RelDC Framework Raw Data Extraction Data Cleaning B Representation Analysis C A D X E Y F ARG Tables/ARGs RelDC Framework Relationship-based Data Cleaning f1 ? f1 f2 ? f2 f3 ? f3 X f4 ? Traditional Methods f4 features and context ARG Y + B C A D X E Y F Relationship Analysis 7 Object Consolidation Notation – O={o1,...,o|O|} set of entities – unknown in general – X={x1,...,x|X|} set of repres. – d[xi] the entity xi refers to – unknown in general – C[xi] all repres. that refer to d[xi] – “group set” – unknown in general – the goal is to find it for each xi – S[xi] all repres. that can be xi – “consolidation set” – determined by FBS – we assume C[xi] S[xi] 8 Attributed Relational Graph (ARG) ARG in RelDC Nodes – per cluster of representations – per representation (for “tough” cases) Edges – regular – similarity person publication department organization 9 Context Attraction Principle (CAP) Take a guess: Who is “J. Smith” – Jane? – John? Jane Smith J. Smith John Smith Merging a new publication. 10 Questions to Answer 1. Does the CAP principle hold over real datasets? That is, if we consolidate objects based on it, will the quality of consolidation improves? 2. Can we design a generic solution to exploiting relationships for disambiguation? 11 Consolidation Algorithm 1. Construct ARG and identify all VCS’s – use FBS in constructing the ARG 2. Choose a VCS and compute c(u,v)’s – for each pair of repr. connected via a similarity edge 3. Partition VSC – – – – use a graph partitioning algorithm partitioning is based on c(u,v)’s after partitioning, adjust ARG accordingly go to Step 2, if more VCS exists 12 Connection Strength Computation of c(u,v) Phase 1: Discover connections – all L-short simple paths between u and v – bottleneck – optimizations, not in IQIS’05 Phase 2: Measure the strength – in the discovered connections – many c(u,v) models exist – we use model similar to diffusion kernels B C A u v D E G F H z 13 Existing c(u,v) Models Models for c(u,v) B – many exists A – diffusion kernels, random walks, etc – none is fully adequate – cannot learn similarity from data C u v D E F Diffusion kernels – (x,y)= 1(x,y) “base similarity” G H z – via direct links (of size 1) – k(x,y) “indirect similarity” – via links of size k – B: where Bxy = B1xy = 1(x,y) – base similarity matrix – Bk: indirect similarity matrix – K: total similarity matrix, or “kernel” 14 Our c(u,v) Model Our model & Diff. kernels N-2 ... ... ... ... ... – virtually identical, but... – we do not compute the whole matrix K MIT T2 – we compute one c(u,v) at a time T1 John Smith T2 T1 P1 Alan White – we limit path lengths by L – (x,y) is unknown in general – the analyst assigns them – learn from data (ongoing work) (a) R1:John (b) R3:John (c) A6:Tom P1 A4:Alan A1:John P4 R3:John P2 MIT A5:Mike P3 R2:J.Smith A4:Alan P1 R1:John Stanford A7:Kate A3:John Our c(u,v) model – regular edges have types T1,...,Tn – types T1,...,Tn have weights w1,...,wn – (x,y) = wi – get the type of a given edge – assign this weigh as base similarity – paths with similarity edges – might not exist, use heuristics 15 Consolidation via Partitioning Observations – each VCS contains representations of at least 1 object – if a repr. is in VCS, then the rest of repr. of the same object are in it too Partitioning 1 1 3 1 2 1 3 2 2 VCS 1 – two cases – k, the number of entities in VSC, is known – k is unknown 4 5 4 – when k is known, use any partit. algo – maximize inside-con, minimize outside-con. – we use [Shi,Malik’2000] – normalized cut 5 5 5 VCS 2 – when k is unknown – split into two: just to see the cut – compare cut against threshold – decide “to split” or “not to split” actually 16 Measuring Quality of Outcome Existing measures – dispersion [DMKD’04] – for an entity, into how many clusters its repr. are clustered, ideal is 1 Ideal Clustering 1 1 1 1 1 1 2 2 2 2 2 2 C1 C2 Div H 1 0 1 0 E1 E2 Dis H 1 0 1 0 – diversity – for a cluster, how many distinct entities it covers, ideal is 1 – easy, clear semantics – but have problems, see figure Entropy – for an entity, if out of m represent. m1 to C1; ...; mn to Cn then One Misassigned (Example 1) 1 1 1 1 1 2 2 2 2 2 2 1 C1 C2 Half Misassigned 1 1 1 2 2 2 2 2 2 1 1 1 C1 C2 H 2 0.65 2 0.65 Div H 2 1 2 1 E1 E2 E1 E2 Dis H 2 0.65 2 0.65 Dis H 2 1 2 1 Dis/Div cannot distinguish the two cases Entropy can: since 0.65 < 1, first clustering is better One Misassigned (Example 2) 1 – if a cluster consists of represent.: m1 of E1; ...; mn of En then (same...) – ideal entropy is zero Div 1 2 1 2 1 1 2 1 2 2 2 C1 C2 Div H 2 0.592 1 0 Dis E1 E2 H 1 0 2 0.65 Average entropy decreases (improves), compared to Example 1 17 Experimental Setup RealMov – movies (12K) – people (22K) – actors – directors – producers – studious (1K) – producing – distributing Uncertainty – d1,d2,...,dn are director entities – pick a fraction d1,d2,...,d10 – group, e.g. in groups of two – {d1,d2}, ... ,{d9,d10} – make all representations of d1,d2 indiscernible by FBS, ... Parameters Baseline 1 – L-short simple paths, L = 7 – L is the path-length limit – one cluster per VCS, regardless – dumb? ... but ideal disp & H(E) Note Baseline 2 – The algorithm is applied to “tough cases”, after FBS already has successfully consolidated many entries! – knows grouping statistics – guesses #ent in VCS – random assigns repr. to clusters 18 Sample Movies Data 19 The Effect of L on Quality Cluster Entropy & Diversity Entity Entropy & Dispersion 20 Effect of Threshold and Scalability 21 Summary RelDC – developed in Aug 2003 (reference disambiguation) – domain-independent data cleaning framework – uses relationships for data cleaning – reference disambiguation [SDM’05] – object consolidation [IQIS’05] Ongoing work – “learning” the importance of relationships from data 22 Contact Information RelDC project www.ics.uci.edu/~dvk/RelDC www.itr-rescue.org (RESCUE) Zhaoqi Chen chenz@ics.uci.edu Dmitri V. Kalashnikov www.ics.uci.edu/~dvk dvk@ics.uci.edu Sharad Mehrotra www.ics.uci.edu/~sharad sharad@ics.uci.edu 23