As data is produced at an unprecedented rate, the need and expectation to make it easily available for the end-users is growing. Dataset Discovery has become an important subject in the data management community, as it represents the means of providing the data to the user and fulfilling an information need. Since the end-user is the one that needs the information and knows what type of information to look for, little has been done to involve the user in the discovery process. This PhD project addresses the topic of interactive data discovery, where the user's interests are modelled through interactions and used as a context for the discovery process. We aim to develop a system that addresses the problem of minimising the trade-of between eficiency and efectiveness, thus providing accurate results in an interactive fashion. The innovative part of the system consists of extracting the user's interests and data needs through interactions and using them to enrich the data context and provide tailored results to the user. We describe the steps to create models and methods that would be used in designing the prototype and we relate to previous systems and neighbouring communities for optimising the system.
It is often said that ”data is the new oil“. We have the means to
produce and collect data at a fast pace and new repositories emerge
such as the data lakes. Having this vast amount of data, both expert
and non-expert users expect to create accurate insights much easier
and faster [
Data discovery is the process of finding relevant data among
thousands of disparate heterogeneous datasets. The relevance is
defined by the users, it is based on their needs and it often means
ifnding joinable or unionable datasets [
Interactive Data Exploration is another research area where the
users actively engage with the system. Interactive data exploration
is the process of extracting knowledge from data when the users
explore the data space without a specific target. Thus, the users
engage in an iterative process where they pose queries, review the
result, adjust the query and repeat the cycle until they decide to
terminate the process [
The main diference between the two processes, the discovery and the exploration, is the user’s data need. In the discovery process, the users have an information need, but are unaware of the location of the data. In the exploration process, they are unaware of both the kind of information that resides in the repository and the kind of information that they need.
We focus on data discovery in data lakes and therefore we
assume that the users have an information need. Providing them
with relevant datasets implies that the data lake contains suficient
information about the datasets such as comments and
documentation about the entities, their semantic types, and the relations
among datasets. However, data lakes only provide minimum
metadata about the datasets, thus they lack the contextual information,
which aids in finding related datasets. In the rest of the paper we
use context to refer to the contextual information, which means
additional data about the datasets. The works focused on the web
tables use as context the text surrounding the tables to infer the
topic and determine which tables match [
However, involving the user raises a new set of challenges. Firstly,
we have to take into account the interaction time and the fact that
the users lose their patience and interest after one second [
Attribute
Recommendation
List of
tables
• We aim to devise a strategy to create easy, meaningful and
suficient interactions to support both the user and the
system in the discovery process.
• We will research ranking mechanisms based on a rich graph
structure containing the key-foreign key constraints,
similarities between attributes and datasets, the local relevance of a
column (being part of a table) as well as the user interactions.
• We will research a model to link the data in order to save the
connections created during a session for the future users.
• We aim to achieve interactive speeds (execution time per
iteration smaller than 1000ms [
Following, we present the research goal in Section 2 and the research overview in Section 3. We describe the challenges we meet to achieve the goal and the shortcoming of existing methods in Section 4 and conclude in Section 5. 2
The goal of this PhD research is to improve the usability and performance of the data discovery process. On one hand, we aim to improve the performance by achieving high accuracy while decreasing the latency and the user efort. On the other hand, we aim to improve the usability by providing the exact result to the users, instead of a list with the top-k results as it is often done in the literature.
Currently, the data discovery systems are faced with the trade-of
between efectiveness and eficiency. Moreover, they return uniform
impersonal results, disregarding the users’ needs and ofering a
limited involvement in the process. These systems focus on automating
the process and minimising the user input as much as possible [
We propose an interactive system, where the users become firstclass citizens. They are in control of the discovery process and by interacting with the system they provide meaningful information about their interests. This information is further modelled and leveraged by the system in order to retrieve the most relevant datasets. We aim to combine these datasets through user interactions and to return to the user one single relevant table. 3
We envision a system that helps the users create tables with N attributes based on keywords. The exploration is clearly defined and it is tailored to assist the user in the process in an iterative manner. The system uses the existing information from the data lake to create the desired result and presents the most relevant datasets to the user based on a number of specific interactions. Through these interactions, the users find and combine the attributes in a single table. We propose an iterative process, where the user and the system engage in a conversation as illustrated in Figure 1.
The user starts the exploration with a keyword and waits for the systems to provide similar attribute names existing in the repository. Next, the user selects one attribute from the list of recommendations and the system starts searching for join candidates. The user receives a list of join candidates and selects a table together with the attributes that might help in the next search. Finally, the process continues with the same steps until the user explicitly stops it. Once the process terminates, the system links the data and saves the interactions for future users.
During this process, the focus is on extracting information
through interactions, illustrated in Figure 1 by the system
monitoring the user. Furthermore, we focus on modelling this
information such that the system recommends meaningful attributes and
datasets to the user. Next, we want to capture the interactions and
use mechanism for short-term memory to understand the user’s
interests and long-term memory to help other users with similar
interests [
In this section, we present the strategy to achieve our goal and
relate to current solutions from the literature. We present their
limitations, the challenges to build our system and propose solutions
to overcome them.
4.1
Most data discovery systems use similar user interactions, such as
providing a target table to find the relevant datasets [
Although some of these data discovery systems involve the user in the process, they do not use the interactions to develop a user profile. We plan on extracting information from the user interactions and develop a user profile to model the users behaviour and interests. With this, we aim to provide more accurate and informative results for the user. However, the human-in-the-loop approach presents a number of challenges: • What is the minimum number of iterations that ofers sufifcient information for the system while keeping the user engaged in the process? • What kind of information can we extract from the interactions in order to model the user’s interests?
To address these challenges we plan on experimenting with
techniques from information retrieval regarding the user search
behaviour [
The efectiveness of an algorithm can be summarised as the
approach used to retrieve the most relevant datasets. One such popular
approach is the top-k ranking. A ranking function replaces the need
of a threshold and eases the discovery process, as the users do not
need prior knowledge to indicate the number of results. In data
discovery, top-k ranking is used in combination with intersection
estimation [
These approaches use the similarities between the datasets under
diferent signals in order to retrieve joinable or unionable datasets.
However, the similarity is highly dependent on the quality of the
data, as shown in our previous work [
The systems using web tables overcome this problem by giving
more context to the data, such as the text surrounding the tables,
the title of the web page, the URL [
Our hypothesis is that a graph structure will allow us to model both the context and the signals as depicted in Figure 2. By context we refer to the user’s interests, while by signals we relate to the data profiles and similarities between the datasets. We will employ several machine learning algorithms both linear and based on decision trees to achieve a ranking function with a high accuracy based on the the context and signals. 4.3
The purpose of data discovery is to facilitate the access to the
data for the integration tasks, such as schema matching, merging,
mapping and query reformulation [
We envision one more step, where the semantic relations are captured, thus working on creating a mapping. As depicted in Figure 1, the last step in the workflow consists of saving all the interactions and linking the data. The challenge consists of creating the proper signals such that we enrich the data structure created in Section 4.2. This new information should help the next users in their discovery process when the system lacks information about their interests (in the beginning of the process, before building the user profile).
Solving the challenges from Section 4.1 and Section 4.2 should point us in the right direction for extracting and modelling the signals to create a mapping and integrate it in the data model used for ranking. 4.4
In the most basic approach, the data discovery is an ( 2) problem,
where every candidate is matched against another. However, the
execution time of the data discovery algorithms becomes a concern
in a large repository that contains thousands or millions of tuples.
Therefore, various approaches have been proposed to minimise the
sample space. Indexing is one of the preferred methods, such as
using LSH [
RESTAU BOOKING User 1: Not similar
SYNTACTIC SIMILAR date name cid PK-FK
RESTAURANT name longitude
HOTEL BOOKING
CITY
cid HOTEL name longitude Nodes to connect Negative edge Positive edge
name { Semantic type: city,
% of nulls: 20, # of distinct values: 30,
Avg length: 8,
Value distribution: [. .] }
with probabilities [
We will follow an empirical methodology to assess what method or which combination of methods helps us create an interactive and efective system. Moreover, we will take into account the power of distributed systems and we envision a hybrid approach consisting of sample space minimisation and distributed algorithms. 5
In conclusion, this paper presents the plan for a PhD project that aims to improve the data discovery process by using the user interactions. We introduce the concept of Interactive Data Discovery and propose a system to model this process. We address several challenges that concern the user interactions and the methodology to achieve high accuracy and low latency. Finally, we are interested in minimising the trade-of between the eficiency and efectiveness by leveraging the user interactions without a significant increase in the user efort.
This work has been partially funded by the H2020 project OpertusMundi No. 870228.