Data exploration promotes a new querying philosophy that gradually converges into queries that can be used to exploit raw data collections according to data explorers (i.e., users) expectations. Data exploration aims to guide the understanding of data collections with diferent rawness degree and define the type of questions that can be asked on top of them, often through interactive exploration processes. This paper introduces a human-guided data exploration approach defining exploration operations that result in diferent types of factual and analytic queries. Our first results include a proposal of query morphing and queries as answers strategies. This paper describes an experiment setting used for testing the data exploration techniques.
The evolution in querying, information retrieval and
humancomputer interaction has led to the shift of interest from
traditional Query-Response paradigm to actual human intelligence
systems. Approaches such as interactive query expansion (IQE)
[
Traditional data management systems assume that when users
ask a query (i) they have good knowledge of the schema,
meaning and contents of the database, and (ii) they are sure that this
particular query is the one they wanted to ask. In short, it is
assumed that users know what they are looking for. In response,
systems like DBMS, always try to produce correct and complete
results [
This paper proposes a data exploration approach that • defines a loop where, given a user intention expressed using terms and raw data collections, the exploration strategies propose diferent types of possible queries that can be asked on top of data, and that potentially correspond to user’s expectations; • interacts with the user for refining intentions based on the proposed queries and starts the loop again until the queries proposal converges with user expectations.
Diferent data exploration pipelines can be defined combining diferent exploration techniques for performing specific data exploration tasks.
The rest of the paper is organised as follows. Section 2 summarises the related work done in the field of data exploration proposing a classification of approaches and techniques. Section 3 provides a detailed description of the approach. Section 4 describes our experimental setting including the dataset and its pre-processing steps. Section 5 concludes the paper and discusses future work. 2
Data exploration calls for combining diferent exploration, querying and processing methods and strategies proposed in diverse domains. Therefore we performed a systematic review to classify them (see Figure 1). According to a systematic review we performed, we propose a classification of existing data exploration techniques and methods. The classification consists of facets representing an aspect of data exploration and dimensions that denote the concepts that define each facet. As shown in Figure 1, the facets classify: (F1) the type of queries addressed by existing work; (F2) the type of algorithms used for exploring data collections; (F3) the knowledge domain of data collections and data types; (F4) the exploration processes done with human intervention; and (F5) data exploration techniques and systems conceived for understanding raw datasets content.
Since exploration can put diferent types of queries in action, facet F1 classifies the types of queries that are defined and used in diferent works that exploit datasets. The spectrum goes from "classic" keyword and relational queries evaluated on top of more or less curated datasets, to data processing operations on raw datasets (e.g., descriptive statistics). In this spectrum, these types represent families of queries that can include aggregation, clustering operations. We mainly identify "query by example" techniques useful particularly in cases where the knowledge about the datasets’ content is too weak (see d1.8). Query by example is an intuitive way to explore data, so many techniques are applying it to data exploration. Examples can either represent approaches like reverse engineering querying and queries like query morphing or queries. We also note that data exploration is a loop that obtains approximated results and the techniques are specialised according to the type of data models (relational, graph, semi-structured, text, multimedia).
Depending on the domain, works propose algorithms rather than operators (like in relational contexts) to process datasets and to discover and derive a precise statistical understanding of their content (facet F2). Algorithms sometimes depend on the type of data structures used for representing data. For example, there are algorithms for processing graphs (centrality, pathfinding, etc.) or querying tables (selection, projection, etc.). Many works use well-known heuristics, data mining, machine learning, artificial intelligence algorithms for processing datasets, and insight into their content. Finally, other works propose their strategies without adhering to a specific domain.
The vision of data exploration in this work is that it should
be a human-guided process. Therefore, we have studied
techniques where humans intervene to adjust and guide the process
of receiving information (d.4.5). We studied works on group
recommendation, consensus functions, group preference and group
disagreement. These study address objectives like designing
consensus functions that aggregate individual group members’
preferences to reflect the overall group’s preference for each item
[
According to our classification, facet F5 considers dimensions that represent exploration techniques. Regarding exploration query expression (d5.1), we have identified three types of approaches: multi-scale query processing for gradual exploration; query morphing to adjust for proximity results; queries as answers as query alternatives to cope with lack of providence. Results filtering (d5.2) addresses analysis and visualisation to give insight to data content. Finally, data exploration systems & environments (d5.3) are tailored for exploring data incrementally and adaptively.
Concerning data exploration techniques, M. L. Kersten et al.
[
Finally, facet F3 classifies the type of datasets used to test different exploration techniques and approaches. Datasets content is often textual and with diferent rawness degree (newspapers, micro-texts from social networks) and already processed content using NLP (Natural Language Processing) techniques and represented as graphs or tables. Other datasets are built by collecting observations monitored using, for example, IoT infrastructures. These data sets contain records of measures or even video or images.
We have the following remarks about what we have studied in state of the art. Data exploration pipelines are mostly ad-hoc, implemented in an artisanal manner, and partially human-guided. Machine learning, analytics and querying techniques (e.g., query by example, queries as answers, etc.) are complementary. We observed that no existing system integrates them so that data scientists can develop exploration pipelines that can thoroughly understand data and its analytics potential. Therefore, there is room for proposing approaches for each of them, defining rules on how they can be combined within data exploration pipelines and integrating them to provide a whole data exploration environment. 3
Figure 2 shows our general approach based on query rewriting techniques and summarised as follows: "given an initial query, provide sets of queries that can help data consumers better exploit data collections". The approach considers that data collections are textual and indexed (not necessarily cleaned) and the representative vocabulary used in their content has been extracted and classified. For example in a crisis management scenario, the classes are events (e.g., someone looks for shelter, a building has been damaged) and actions (e.g., a hotel provides shelter for victims, people is approaching a damaged building to search victims).
The approach is intended to rewrite initial keyword queries by morphing expressions to produce results that can retrieve representative insights into these collections’ content. The rewriting process is gradual and interactive, where the user expresses an initial expression, and the exploration process provides new queries associated with content samples that can give insight into the content of the dataset. The alternative queries are assessed and adjusted by the user. Then, the exploration process is triggered again until a set of queries is chosen to be evaluated to produce results. Results produced by diferent exploration strategies can also be used as input to others. For example, query morphing’s output can be used as an input for the queries as answers.
The next sections describe two rewriting techniques query morphing and queries as answers (expansion) that we have proposed for exploring datasets. 3.1
Query morphing is the process of rewriting conjunctive and disjunctive keyword queries, by adding terms, to increase the possibility of exploring the most number of items in a collection. We proposed and implemented a “query morphing” pipeline that can help the data scientist better precise her query (see Figure 3). Our query morphing pipeline uses a vocabulary and Wordnet to look for associated terms and synonyms that help expand the terms to enhance the chance of matching it with relevant data items in the target collection. The pipeline is described as follows. Given a conjunctive and disjunctive keyword query represented as an expression tree go through the tree in depth-first until ifnding a leaf representing a term and then: (1) Use a vocabulary representing the dataset content and wordnet seeking for: (a) equivalent terms and generate a node with the operator and then connect the initial term with the equivalent terms in a conjunctive expression subtree. (b) more general terms and connect the initial term with these terms in a disjunctive expression subtree. (c) assess and adjust the morphed query by the user and eventually restart the expansion process. The assessment process includes pondering the terms and exploring result samples to see potential results. For example, for a query "victims AND missing AND shelter", using Wordnet1, the query can be expanded as follows: "(victim OR casualty OR "unfortunate person") AND (missing OR absent) AND (shelter OR protection OR housing)". The key is using a concept ontology or glossary and can find the maximum equivalent and general terms. In this example, we only found equivalent terms. The user can then mark which terms should or should not be included in the expanded query. She can also test diferent combinations of the query and compare the results to see which morphed query can produce the results that best respond to her expectations.
Once the new query expression has been rewritten, as done in information retrieval techniques, we use the inverted index to ifnd the corresponding documents where the query terms occur. Then, we use the frequency matrix to compute the final result set tagged with precision and recall measures.
Given an initial conjunctive/disjunctive keyword query, the query is rewritten and transformed into several queries by extending it with general and more specific terms, synonyms, etc., and by exploiting the knowledge domain (see Figure 3). The result is a set of possible alternative queries with associated sample results so that the user can choose which ones to execute.
In our approach, the initial query is represented by an expression tree (intermediate representation) where nodes are conjunction and disjunction operators and leaves are terms. During the rewriting process, the tree is modified by adding new types of nodes and tagged arcs. New nodes represent “and” and “or” nodes that do not belong to the initial query and more general/specific terms associated with an “initial” term. These new nodes are connected with the nodes of the initial query by tagged arcs. A tag can indicate whether it connects a node with a conjunction or disjunction of more general/precise terms. For example, consider an initial query linking three terms t1, t2 and t3 with a conjunctive and disjunctive initial query (see Figure 3). It is then re-written in a new query represented by a tree that extends the query with terms that can be synonyms or related terms to t1 and t2. Three possible queries are derived: Q1 1 which provides an alternative to t1 with a new complex query with a synonym/or an associated term t4, saying that we can look for "t1 or t1 and t4". Similarly, Q12 provides an alternative to t2 saying that we can either look for t2 or t2 and t6, which could be a synonym or a related term. Finally, Q13 is a complex query that integrated Q11, Q12 with t3 of the initial query.
The following steps are performed for computing queries alternatives where every step aims at deriving the initial query into queries that add knowledge. For each leaf in the expression tree of the query: (1) Use a vocabulary (extracted from the dataset content) and Wordnet seeking for: • equivalent terms and generate a node with the operator and then connect the initial term with the equivalent terms in a conjunctive expression subtree; • more general terms and connect the initial term with these terms in a disjunctive expression subtree. (2) Use a frequency matrix for looking for terms that are often associated with the initial term with a specific frequency and getting a sample of documents that can belong to query results.
The user can choose those queries that best target her expectations. A history of queries is maintained that can be reused for suggesting or pre-calculating morphing or query as answers results or for adjusting the chosen query set with new queries as the dataset evolves. 4
To experiment our general approach, let us consider a disaster
management scenario where various data collections are
produced during the life cycle of the disaster and must be explored
to organise relief, resilience and duty of memory actions. The
scenario we use is related to disaster management under a
horizontal organisation 2, where civilians take active action when an
event happens (e.g., earthquake, flooding, fire) and continue to
influence decision making during the other phases of its
management. In this context, social media is a fast-paced channel used by
afected people to describe their situation and observations, seek
information, specify their requests, and ofer their voluntary
assistance; providing actionable information [
During such life-threatening emergencies, afected and vulnerable people, humanitarian organisations, and other concerned authorities search for information useful to provide help and prevent a crisis. Nobody has control over the type of data exchanged by actors. These data are crucial in making critical decisions like saving lives, searching people, and providing shelter and medical assistance. For this reason, it is required to explore past and present in an agile manner to find hints to make decisions and act individually and collectively. Social network data collections can include reports on architectural and building environment damages and volunteers informing them that they have answered calls for help (see Figure 4).
The question is whether these data collections can help to ifnd (i) causal correlations, for example, is it possible to know given a post asking for help whether actions have already been 2The phenomenon of organisation of civilians under horizontal and marginal groups has come up in diferent countries during rivers flooding, annual landslides in diverse regions particularly in Latin America. taken? since when and whether the problem has been solved; (ii) is it possible to find patterns showing which zones have been systematically damaged in other events? Is there more risk and help required in those regions?; (iii) Spatio-temporal relations, is it possible to figure out from the beginning of the event until a given subsequent time? have actors installed camps to provide ifrst aid? Does help come from urban areas?; (iv) how to ask about the type of help still being required after a day of the event?
Note that these questions are not asking for results, they are asking for assistance on how to ask them on top of data to potentially best explore data. How can I express my query to expect to receive the best guidance to act? Is my query pertinent to be asked given the data I can have access to?
Data exploration techniques can help assist in expressing queries that can potentially explore data collections and be pertinent according to their content.
This section describes the experimental setting for the assessment of our approach. Our experiments deal with the crisis scenario introduced previously, and they use micro-texts datasets from Twitter concerning this topic. Given Twitter’s 140 characters limit restriction, the frequency matrix cannot be useful, so we used the word2vec model to pre-process the dataset and then ifnd similar terms for rewriting queries. In this work, we consider those words provided by word2vec model, and words are also indexed in the frequency matrix for extending query. With this information, we modify the tree adding "AND" and "OR" nodes, and thereby we create other possible queries that derive from the initial one.
We have experimented with generating the knowledge domain and then using it for validating morphing queries. Our experiment is based on the disaster management use case using Twitter posts as documents collections. The experiment applies text mining techniques to build the vocabulary and classify it into events produced and actions performed during a disaster life-cycle. In this section, we first describe the datasets we used in our experiments and then the experiment setting, including the algorithms used to process data collections and classify the extracted vocabulary. 4.1
Among social media studies, most of them focus on Twitter,
mainly because of its timeliness and availability of information
from a large user base. We use CrisisNLP [
Data Preprocessing. Since the tweet texts are brief, informal,
noisy, unstructured, and often contain misspellings and
grammatical mistakes, preprocessing must be done before using them
in further analysis. Moreover, due to Twitter’s 140 characters
limit restriction, Twitter users intentionally shorten words
using abbreviations, acronyms, slang, and sometimes words
without spaces; hence we need to normalise those OOV terms [
(4) We removed duplicate tweets.
After performing the cleaning 126161 unlabelled data related to the 2014 California Earthquake, we obtained a set of 73562 tweets. This set is used for all experiments reported in this work.
The pipeline implemented to create the knowledge base required for experimenting our data exploration techniques consists of two steps: (i) indexing data collections content using information retrieval techniques; (ii) create a vocabulary using classification techniques.
Indexing the data collection. As a result of indexing the cleaned tweets collection, we created an inverted index and a frequency matrix representing the content of the collection. We implemented an inverted index to provide agile access to a document’s position in which a term appears. The inverted index is used as a dictionary that associates each word with a list of document identifiers where the word appears. This structure prevents making the running time of token comparisons quadratically. So, instead of comparing, record by record, each token to every other token to see if they match, the inverted indices are used to look up records that match a particular token.
Currently, we use 73562 unlabelled data set related to the 2014 California Earthquake. We generated an inverted index consisting of 20313 rows. The rows correspond to terms in our raw data collection, and columns correspond to documents where the terms occur. The inverted index allows a fast full-text search. It can help to explore queries’ terms to find the documents where the terms occur.
A term frequency matrix is a mathematical matrix that
describes the frequency of terms that occur in a collection of
documents [
Creating a vocabulary. We implemented a classification pipeline
to build a vocabulary of events and actions related to disasters,
thereby generating a knowledge base describing the tweets’ data
collections used for our experiment. The pipeline combines
machine learning methods reproducing an existing work proposed
by [
We applied supervised techniques such as Random Forest [
We performed a set of experiments on California and Nepal earthquake datasets consisting of approximately 3032 labelled tweets, out of which 2203 tweets of Nepal and 829 tweets of California dataset. The datasets are divided into two sets. As usual in machine learning techniques, we divided the data collection into training and test datasets. The first set comprised of 70% of the messages (i.e. training set) and the second comprised of 30% of the messages (i.e. test set). We trained all three diferent kinds of classifiers using the preprocessed data.
We used multilayer perceptron with a CNN. We conducted experiments on the same dataset and eventually established that CNN outperformed the task with an adequate margin compared to our previous work.
For the evaluation of the trained models, we compared the
results to [
For assessing expanded term quality, we have compared the performance of our proposed classification based query expanding method against the traditional query expanding method. We calculated the mean average of Cosine Similarity (MACS) between the query and expanded query terms to assess the proposed approach’s performance. The experimental results show that the expanded query terms, obtained from the classified query expansion model, are more similar and relevant than the nonclassification model.
In this section, we presented an ablation study about the
performance of our proposed classification based query morphing
method. We use the available crisisNLP pre-trained word
embedding via word2vec method [
( , ) =
Í
=1 ( (_, _ [])) (1)
where CTs are candidate terms, ’m’ is a hyper-parameter in
query expansion-based retrieval, which shows the number of
expansion terms (ET), using as reference the studies[
Currently, we used pseudo relevance feedback. This method automates the manual part of relevance feedback. It is assumed that the user takes top-m ranked morphed query terms returned by the initial query as relevant to expand her query. Results scoring must be completed with user feedback that finally guides the process. We have proposed a solution for exploring scientific papers through an experiment defining a set of exploration queries. Results were assessed by scientists of the National Institute of Genetic Engineering and Biotechnology, Tehran, Iran and Golestan University of Medical Sciences. Scientists provided feedback about exploration operations through questionnaires that are processed for obtaining satisfaction metrics. We are currently defining a crowd-based setting for obtaining feedback in the case of crisis datasets. The idea is to work with diferent groups of users (victims, volunteers, logistics decision-makers, police, medical staf) and queries to assess exploration results. 5
This paper introduced a general datasets exploration approach that includes human in the loop. The current approach includes two exploration techniques (i.e., query morphing and queries as answers) to help define queries that can fully explore and exploit a dataset. They are complementary query rewriting techniques where initially expanding a query can help adjust the terms used for exploring a dataset and then produce possible combinations of terms with possible queries that can lead to diferent scopes. In both cases, the user finally chooses a set of representative queries to her interests and the produced results that target her expectations. We have tested query morphing in the case of crisis dataset exploration, where people involved in a critical event either as victim or volunteers can define queries for retrieving information to look for or provide help.
Our future work includes modelling query exploration pipelines that can combine diferent techniques for exploring data collections. We will also propose ways of morphing and giving queries as answers where queries can be analytical or imply quantitative data views.