Continuous Benchmark of Filtering methods for Entity Resolution

Entity Resolution constitutes a core data integration task of quadratic time complexity. As a result, it scales to large datasets through methods that filtering techniques, which reduce its computational cost to candidate pairs. These are pairs of similar entity profiles that are highly likely to be matching. Three types of filtering methods have been proposed in the literature:

1. Blocking workflows
2. String similarity joins
3. Nearest neighbor methods

The goal of this repository is to provide code and data for benchmarking the relative performance of the main methods of these three types.

Detailed results of our extended experimental analysis are provided here.

We plan to keep our benchmark up-to-date, including new filtering methods and datasets shortly after their publication. Please create a new issue in order to request the addition of a new dataset or filtering method.

Blocking Workflows

Each blocking workflow consists of four steps:

1. Block Building, which is a mandatory step that produces a set of blocks. The following methods have been considered so far:
   1. Standard Blocking
   2. Q-Grams Blocking
   3. Extended Q-Grams Blocking
   4. Suffix Arrays Blocking
   5. Extended Suffix Arrays Blocking
2. Blocking Purging, which is an optional, parameter-free step that discards blocks with more than half the input entities.
3. Block Filtering, which is an optional step that retains every entity in r% of its smallest blocks.
4. Comparisonn Cleaning, which is a mandatory step that applies one of the following methods:
   1. Comparison Propagation, which simply cleans a set of overlapping blocks from all repeated candidate pairs.
   2. Meta-blocking, which assigns a score to every candidate pairs that is proportional to its matching likelihoood and thenn discards the lowest weighted pairs. This way, it removes all repeated candidate pairs as well as a large portion of the non-matching ones. Meta-blocking consists of two parts:
      1. A weighting scheme, the scoring function
      2. A pruning algorithm, which cleans the candidate pairs

In our experiments, we thoroughly fine-tune the configuration parameters of these 5 workflows. We also consider 2 baseline workflows:

1. Parameter-free Blocking Workflow, which combines the three parameter-free methods, i.e., Standard Blocking, Block Purging and Comparison Propagation.
2. Default Q-Grams Blocking Workflow, which combines Q-Grams Blocking with Block Filtering and Meta-blocking, using the configuration parameters determined in a past experimental analysis.

All code and data are available here.

String similarity joins

The following state-of-the-art string similarity join algorithms are considered:

1. ε-Join
2. kNN-Join

In our experiments, we thoroughly fine-tune the configuration parameters of these five workflows. We also consider one baseline method:

1. Default kNN-Join

All code and data are available here.

Nearest neighbor methods

The following state-of-the-art NN methods are considered:

1. MinHash LSH
2. Crosspolytope LSH
3. Hyperplane LSH
4. FAISS
5. SCANN
6. DeepBlocker

In our experiments, we thoroughly fine-tune the configuration parameters of these 6 methods. We also consider one baseline method:

1. Default DeepBlocker (DDB)

All code and data are available here.

Datasets

For the time being, the following real-world datasets for Clean-Clean ER have been used in our experimental study:

Dataset Name	D1 Entities	D2 Entities	D1 Name-Value Pairs	D2 Name-Value Pairs	Duplicates	Average NVP per Entity	Brute-force Comparisons
D1 (Restaurants)	339	2,256	1,130	7,519	89	3.3	7.64E+05
D2 (Abt-Buy)	1,076	1,076	2,568	2,308	1,076	2.4	1.16E+06
D3 (Amazon-Google Products)	1,354	3,039	5,302	9,110	1,104	3.9	4.11E+06
D4 (DBLP-ACM)	2,616	2,294	10,464	9,162	2,224	4.0	6.00E+06
D5 (IMDB-TMDB)	5,118	6,056	21,294	23,761	1,968	4.0	3.10E+07
D6 (IMDB-TVDB)	5,118	7,810	21,294	20,902	1,072	3.2	4.00E+07
D7 (TMDB-TVDB)	6,056	7,810	23,761	20,902	1,095	2.2	4.73E+07
D8 (Amazon-Walmart)	2,554	22,074	14,143	114,315	853	5.2	5.64E+07
D9 (DBLP-Scholar)	2,516	61,353	10,064	198,001	2,308	4.0	1.54E+08
D10 (Movies)	27,615	23,182	155,436	816,009	22,863	5.6	6.40E+08

We have also included the following synthetic datasets for Dirty ER in the scalability analysis of our experimental study:

Dataset Name	Entities	Name-Value Pairs	Duplicates	Average NVP per Entity	Brute-force Comparisons
D1OK	10,000	106,108	8,705	10.61	5.00E+07
D50K	50,000	530,854	43,071	10.62	1.25E+09
D100K	100,000	1,061,421	85,497	10.61	5.00E+09
D200K	200,000	2,123,728	172,403	10.62	2.00E+10
D300K	300,000	3,184,885	257,034	10.62	4.50E+10
D1M	1,000,000	10,617,729	857,538	10.62	5.00E+11
D2M	2,000,000	21,238,252	1,716,102	10.62	2.00E+12

Technical report

More details are provided in the following technical report:

George Papadakis, Marco Fisichella, Franziska Schoger, George Mandilaras, Nikolaus Augsten, Wolfgang Nejdl: How to reduce the search space of Entity Resolution: with Blocking or Nearest Neighbor search?" (pdf).