Internet users have personal data spread over several devices and across several web systems. In this paper, we introduce a novel open-source framework for integrating the data of a user from diferent sources into a single knowledge base. Our framework integrates data of diferent kinds into a coherent whole, starting with email messages, calendar, contacts, and location history. We show how event periods in the user's location data can be detected and how they can be aligned with events from the calendar. This allows users to query their personal information within and across diferent dimensions, and to perform analytics over their emails, events, and locations. Our system models data using RDF, extending the schema.org vocabulary and providing a SPARQL interface.
Internet users commonly have their personal data spread over several devices and services. This includes emails, messages, contact lists, calendars, location histories, and many other. However, commercial systems often function as data traps, where it is easy to check information in but dificult to query and exploit it. For example, a user may have all her emails stored with an email provider – but cannot find out which of her colleagues she interacts most frequently with. She may have all her location history on her phone – but cannot find out which of her friends’ places she spends the most time at. Thus, a user often has paradoxically no means to make full use of data that she has created or provided. As more and more of our lives happen in the digital sphere, users are actually giving away part of their life to external data services.
We aim to put the user back in control of her own data. We introduce a novel framework that integrates and enriches personal information from diferent sources into a single knowledge base (KB) that lives on the user’s machine, a machine she controls. Our system, Thymeflow, replicates data of diferent kinds from outside services and thus acts as a digital home for personal data. This provides the user with a high-level global view of that data, which she can use for querying and analysis. All of this integration and analysis happens locally on the user’s computer, thus guaranteeing her privacy.
Designing such a personal KB is not easy: Data of completely diferent nature has to be modeled in a uniform manner, pulled into the knowledge base, and integrated with other data. For example, we have to find out that the same person appears with diferent email addresses in address books from diferent sources. Standard KB alignment algorithms do not perform well in our scenario, as we show in our experiments. Furthermore, integration spans data of diferent modalities: to create a coherent user experience, we need to align calendar events (temporal information) with the user’s location history (spatiotemporal) and place names (spatial).
We provide a fully functional and open-source personal
knowledge management system. A first contribution of our work is the
management of location data. Such information is becoming
commonly available through the use of mobile applications such as
Google’s Location History [
A second contribution is the adaptation of ontology alignment techniques to the context of personal KBs. The alignment of persons and organizations is rather standard. More novel are alignments based on time (a meeting in the calendar and a GPS location), or space (an address in contacts and a GPS location).
Our third contribution is an architecture that allows the integration of heterogeneous personal data sources into a coherent whole. This includes the design of incremental synchronization, where a change in a data source triggers the loading and treatment of just these changes in the central KB. Conversely, the user is able to perform updates on the KB, which are made persistent wherever possible in the sources. We also show how to integrate knowledge enrichment components into this process, such as entity resolution and spatio-temporal alignments.
As implemented, our system can provide answers to questions such as: Who have I contacted the most in the past month (requires alignments of diferent email addresses)? How many times did I go to Alice’s place last year (requires alignment between contact list and location history)? Where did I have lunch with Alice last week (requires alignment between calendar and location history)?
Our system, Thymeflow, was previously demonstrated in [
We first introduce out data model and sources in Section 2, and
then present the system architecture of Thymeflow in Section 3.
1https://github.com/thymeflow/thymeflow
Section 4 details our knowledge enrichment processes, and Section 5
our experimental results. Related work is described in Section 6.
Before concluding in Section 8, we discuss lessons learnt while
building and experimenting with Thymeflow in Section 7.
to associate it to geo-coordinates and richer place semantics. The
Facebook Graph API [
In this section, we briefly describe the schema of the knowledge base, and discuss the mapping of data sources to that schema.
Schema. We use the RDF standard [
For modeling personal information, we use the schema.org vocabulary when possible. This vocabulary is supported by Google, Microsoft, Yahoo, and Yandex, and documented online. Wherever this vocabulary is not fine-grained enough for our purposes, we complement it with our own vocabulary, that lives in the namespace http://thymeflow.com/personal# with prefix personal.
Figure 1 illustrates a part of our schema. Nodes represent classes, rounded colored ones are non-literal classes, and an edge with label p from X to Y means that the predicate p links instances of X to instances of type Y . We use locations, people, organizations, and events from schema.org, and complement them with more fine-grained types such as Stay, EmailAddress, and PhoneNumber. Person and Organization classes are aggregated into a personal:Agent class.
Emails and contacts. We treat emails in the RFC 822 format [
Email addresses are great sources of knowledge. An email address such as “jane.doe@inria.fr” provides the given and family names of a person, as well as her afiliation. However, some email addresses provide less knowledge and some almost none, e.g., “j4569@gmail.com”. Sometimes, email fields contain a name, as in “Jane Doe <j4569@gmail.com>”, which gives us a name triple. In our model, personal:Agent instances extracted from emails with the same combination of email address and name are considered indistinguishable (i.e., they are represented by the same URI). An email address does not necessarily belong to an individual; it can also belong to an organization, as in edbt-school-2013@imag.fr or fancy_pizza@gmail.com. This is why, for instance, the sender, in our data model, is a personal:Agent, and not a schema:Person.
A vCard contact [
Calendar. The iCalendar format [
A personal knowledge base could be seen as a view defined over
personal information sources. The user would query this view in a
mediation style [
Thymeflow is a web application that the user installs, providing it with a list of data sources, together with credentials to access them (such as tokens or passwords). The system accesses the data sources and pulls in the data. All code runs locally on the user’s machine. None of the data leaves the user’s computer. Thus, the user remains in complete control of her data. The system uses adapters to access the sources, and to transform the data into RDF. We store the data in a persistent triple store, which the user can query using SPARQL.
One of the main challenges in the creation of a personal KB is the temporal factor: data sources may change, and these updates should be reflected in the KB. Changes can happen during the initial load time, while the system is asleep, or after some inferences have already been computed. To address these dynamics, Thymeflow uses software modules called synchronizers and enrichers. Figure 2 shows synchronizers on the left, and enrichers in the center. Synchronizers are responsible for accessing data sources, enrichers (see Section 4) for inferring new statements, such as alignments between entities obtained by entity resolution.
Modules are scheduled dynamically and may be triggered by
updates in the data sources (e.g., calendar entries) or by new pieces
of information derived in the KB (e.g., the alignment of a position in
the location history with a calendar event). The modules may also be
started regularly for particularly costly alignment processes. When
a synchronizer detects a change in a source, a pipeline of enricher
modules is triggered, as shown in Figure 2. Enrichers can also
use knowledge from external data sources, such as Wikidata [
Synchronizer modules are responsible for retrieving new data from a data source. For each data source that has been updated, the adapter for that particular source transforms the source updates since last synchronization into a set of insertions/deletions in RDF. dateTime
dateReceived/ dateSent esntadrDtDataete/ text/ headline string
GeoCoordinates item geo geo l o c a t i o eo n g Place addressLocality/ addressRegion address inReplyTo recipsieennder/
t atendee/ organizer name/ description homweoLrokcLaotcioanti/on fliiitaaon location postalCode/ streetAddress/ postOficeBoxNumber PostalAddress addressCountry time itagnm l/eagn GeoVector velocity u d uncertainty e double longitude/
latitude string Legend P1 Pn X1 Xm
l i a m e
e on h telep name
Address name string e/ givenNilyaNmame fam birthD ate dateTime
address X p schema:X schema:p
X p personal:X personal:p
X xsd:X rdfs:subClassOf This is of course relatively simple for data sources that track modiifcations, e.g., CalDAV (calendar), CardDAV (contacts) and IMAP (email). For others, this requires more processing. The result of this process is a delta update, i.e., a set of updates to the KB since the last time that particular source was considered.
The KB records the provenance of each newly obtained piece of information. Synchronizers record a description of the data source, and enrichers record their own name. We use named graphs to store the provenance. For example, the statements extracted from an email message in the user’s email server will be contained in a graph named with the concatenation of the server’s email folder URL and the message id. The graph’s URI is itself an instance of personal:Document, and is related to its source via the personal:documentOf property. The source is an instance of personal:Source and is in this case the email server’s URL. Account information is included in an instance of personal:Account via the personal:sourceOf property. Account instances allows us to gather diferent kinds of data sources, (e.g., CardDAV, CalDAV and IMAP servers) belonging to one provider (e.g., corporate IT services) to which the user accesses through one identification. This provenance can be used to answer queries such as “What meetings were recorded in my work calendar for next Monday?”.
Finally, the system allows the propagation of information from the KB to the data sources. These can either be insertions/deletions derived by the enrichers, or insertions/deletions explicitly specified by the user. For instance, consider the information that diferent email addresses correspond to the same person. This information can be pushed to data sources, which may for example result in performing the merge of two contacts in the user’s list of contacts. To propagate the information to the source, we translate from the structure and terminology of the KB back to that of the data source and use the API of that source. The user has the means of controlling this propagation, e.g., specifying whether contact information in our system should be synchronized to her phone’s contact list.
The user can update directly the KB by inserting or deleting
knowledge statements. Such updates to the KB are specified in the
SPARQL Update language [
We describe the general principles of enricher modules. We then describe two specific enrichments: agent matching and event geolocation.
After loading, enricher modules perform inference tasks such as entity resolution, event geolocation, and other knowledge enrichment tasks. An enricher works in a diferential manner: it takes as input the current state of the KB, and a collection of changes ∆ i that have recently happened. It computes a new collection ∆ i+1 of enrichments. Intuitively, this allows reacting to changes in a data source. When some ∆ 0 is detected (typically by some synchronizer), the system runs a pipeline of enrichers to take these changes into consideration. For instance, when a new entry is entered in the calendar with an address, a geocoding enricher is called to locate it. Another enricher will later attempt to match it with a position in the location history. For performance, particularly costly enrichers wait until there are enough changes, or when no more changes are happening, before running on a batch of changes. This is the case for the entity resolution enricher. We now present this enricher and another one that has been incorporated into the system. 4.1
Facets. The KB keeps information as close to the original data as
possible. Thus, the knowledge base will typically contain several
entities for the same person, if that person appears with diferent
names or diferent email addresses. We call such resources facets of
the same real-world agent. Diferent facets of the same agent will
be linked by the personal:sameAs relation. The task of identifying
equivalent facets has been intensively studied under diferent names
such as record linkage, entity resolution, or object matching [
Identifier-based matching. We can match two facets if they have the same value for some particular attribute (such as an email address or a telephone number), which, in some sense, identifies or determines the entity. This approach is commonly used in personal information systems (in research and industry) and gives fairly good results for linking, e.g., facets extracted from emails and the ones extracted from contacts. Such a matching may occasionally be incorrect, e.g., when two spouses share a mobile phone or two employees share the same customer relations email address. In our experience, such cases are rare, and we postpone their study to future work.
Two agent facets with the same first and family names have, for instance, a higher probability to represent the same agent than two agent facets with diferent names, all other attributes held constant. Besides names, attributes that can help determine a matching include schema:birthDate, schema:gender, and schema:email.
We tried holistic matching algorithms for graph alignments [
We discuss how to geolocate events, e.g., how we can detect that Monday’s lunch was at “Shana Thai Restaurant, 311 Mofett Boulevard, Mountain View, CA 94043”. For this, we first analyze the location history from the user’s smartphone to detect places where the user stayed for a prolonged period of time. We then perform some spatiotemporal alignment between such stays and the events in the user’s calendar. Finally, we use geocoding to provide location semantics to the events, e.g., a restaurant name and a street address.
Detecting stays. Locations in the user’s location history can be put into two categories: stays and moves. Stays are locations where the user remained for some period of time (e.g., dinner at a restaurant, gym training, ofice work), and moves are the others. Moves usually correspond to locations along a journey from one place to another, but might also correspond to richer outdoor activity (e.g., jogging, sightseeing). Figure 3 illustrates two stay clusters located inside the same building.
To transform the user’s location history into a sequence of stays
and moves, we perform time-based spatial clustering [
Q P d(µ P , µ Q ) is the geographical distance between cluster centers. The Hellinger distance takes into account both the accuracy and geographical distance between cluster centers, which allows us to handle outliers no matter the location accuracy. The location is added to C if this distance is below a certain threshold λ, i.e., H 2(P , Q) ⩽ λ2 < 1. In our system, we used a threshold of 0.95.
When p is added to cluster C, the resulting cluster is defined with a normal distribution whose expectation is the arithmetic mean of location point centers weighted by the inverse accuracy squared, and whose variance is the harmonic mean of accuracies squared. Formally, if a cluster C is formed by locations {p1, . . . , pn }, where pi = (ti , xi , yi , ai ), then C is defined with distribution N (µ , σ 2) where µ is the weighted arithmetic mean of location centers (xi , yi ) weighted by their inverse accuracy squared ai−2, and the variance σ 2 is the harmonic mean of location accuracies squared ai−2. µ = Xn (xi , yi ) Xn 1 i=1 ai2 i=1 ai2 !−1 σ 2 =
Xn 1
2 i=1 ai !−1 The coordinates are assumed to have been projected to an Euclidean plane locally approximating distances and angles on Earth around cluster points. If n = 1, then µ = (x1, yi ) and σ 2 = a12, which corresponds to the definition of a cluster of size 1.
A cluster that lasted more than a certain threshold is a candidate for being a stay. A dificulty is that a single location history (e.g., Google Location History) may record locations of diferent devices, e.g., a telephone and a tablet. The identity of the device may not be recorded. The algorithm understands that two far-away locations, very close in time, must come from diferent devices. Typically, one of the devices is considered to be stationary, and we try to detect a movement of the other. Another dificulty comes when traveling in high speed trains with poor network connectivity. Location trackers will often give the same location for a few minutes, which leads to the detection of an incorrect stay.
Matching stays with events. After the extraction of stays using the previous algorithm, the next step is to match these with calendar events. Such a matching turns out to be dificult because: (i) the location of an event (address or geo-coordinates) is often missing; (ii) when present, an address often does not identify a geographical entity, as in “John’s home” or “room C110”; (iii) in our experience, starting times are generally reasonable (although a person may be late or early for a meeting) but durations are often not meaningful (around 70% of events in our test datasets were scheduled for 1 hour; among the 1-hour events that we aligned, only 9% lasted between 45 and 75 minutes); (iv) some stays are incorrect.
Because of (i) and (ii), we do not rely much on the location explicitly listed in the user’s calendars. We match a stay with an event primarily based on time: the time overlap (or proximity) and the duration. In particular, we match the stay and the event, if the ratio of the overlap duration over the entire stay duration is greater than a threshold θ . As we have seen, event durations are often unreliable because of (iii). Our method still yields reasonable results, because it tolerates errors on the start of the stay for long stays (because of their duration) and for short ones (because calendar events are scheduled usually for at least one hour). If the event has geographical coordinates, we filter out stays that are too far away from that location (i.e., when the distance is greater than δ ). We discuss the choice of θ and δ for this process in Section 5.
Geocoding event addresses. Once stays associated with events,
we enrich events with rich place semantics (country, street name,
postal code, place name). If an event has an explicit address, we use
a geocoder. Thymeflow allows using diferent geocoders, e.g., the
Google Maps Geocoding API [
In this section, we present the results of our experiments. We used datasets from two real users, whom we call Angela and Barack. Angela’s dataset consists of 7,336 emails, 522 calendar events, 204,870 location points, and 124 contacts extracted from Google’s email, contact, calendar, and location history services. This corresponds to 1.6M triples in our schema. Barack’s dataset consists of 136,301 emails, 3,080 calendar events, 1,229,245 location points, and 582 contacts extracted from the same sources. Barack’s emails cover a period 5,540 days, locations cover 1,676 days. This corresponds to 10.3M triples, where 70.9 % come from the location history, 28.8 % from emails, 0.3 % from calendars and less than 0.1 % from contacts.
We measured the loading times of Angela’s dataset into the system in two diefrent scenarios: from source data on the Internet (using Google API, except for the location history which is not provided by the API and was loaded from a file), and from source data stored in local files. Loading took 19 and 4 minutes, respectively, on a desktop computer (Intel i7-2600k 4-core, 3.4 GHz, 20 GB RAM, SSD). 5.1 We evaluated the precision and recall of the AgentMatch algorithm (Section 4) on Barack’s dataset. This dataset contains 40,483 Agent instances with a total of 25,381 schema:name values, of which 17,706 are distinct; it also contains 40,455 schema:email values, of which 24,650 are distinct. To compute the precision and recall, we sampled 2,000 pairs of distinct Agents, and asked Barack to manually assign to each possible pair a ground truth value (true/false). Barack was provided with the email address and name of each agent, and was allowed to query the KB to get extra information.
We tested both Levenshtein and Jaro–Winkler as secondary string distance, with and without IDF term weights. The term qgram match ratio (S) was set to 0.6. We varied λ so as to maximize the F1 value. Precision decreases while recall increases for decreasing threshold values. Our baseline is IdMatch, which matches two contacts if they have the same email address.
As competitor, we considered PARIS [
We also considered Google’s “Find duplicates” feature. Google was not able to handle more than 27,000 contacts at the same time, and so we had to run it multiple times in batches. Since the final output depends on the order in which contacts were loaded, we present two results, one for which the contacts were supplied sorted by email address (Google1), and another for a random order (Google2). Since Google’s algorithm failed to merge contacts that IdMatch did merge, we also tested running IdMatch on Google’s output (GoogleId) for both runs. We also tested Mac OS X contact de-duplication feature. However, its result did not contain all the meta data from the original contacts, so that we could not evaluate this feature.
The results are shown in Table 1. As expected, our baseline IdMatch has a perfect precision, but a low recall (43%). Google, likewise, gives preference to precision, but achieves a higher recall than the baseline (50%). The recall improves further if the Google is combined with IdMatch (61%). PARIS, in contrast, favors recall (92%) over precision (83%), and achieves a better F1 value overall. The highest F1-measure (95%) is reached for AgentMatch with the Jaro–Winkler distance for a threshold of 0.825. It has a precision comparable to Google’s, and a recall comparable to PARIS’s. 5.2
We evaluated the extraction of stays from the location history on
Barack’s dataset. We randomly chose 15 days, and presented him
a web interface with (1) the raw locations on a map, (2) the stays
detected by Thymeflow, and (3) the stays detected by his Google
Timeline [
Table 2 shows the resulting precision and recall for each method, with varying stay duration thresholds for Thymeflow. We also show the duplicate ratio #D (number of duplicates over the number of true stays) and the move duration ratio DΘ (duplicate move duration over the total duration of true stays). Overall, Thymeflow obtains results comparable to Google Timeline for stay extraction, with better precision and recall, but more duplicates.
Matching stays with events. We also evaluated the matching of stays in the user’s location history with the events in their calendar. We sampled stays from Angela and Barack’s datasets, and produced all possible matchings to events, i.e., all matchings produced by the algorithm whatever the threshold. Angela and Barack were then asked to manually label the matches as correct or incorrect. The matching process relies on two parameters, namely the duration ratio threshold θ and the filtering distance δ . We varied θ and found that a value of 0.2 leads to best F1 values. With this value, we varied F1 δ , and found that the performance improves consistently with larger values. This indicates that filtering stays which are too far from event location coordinates (where available) should not be taken into consideration. With these settings, the matching performs quite well: We achieve a precision and recall of around 70%. 5.3
We evaluated diferent geocoding enrichers for (event, stay) pairs described in Section 4. We used the Google Maps Geocoding API. We considered three diferent inputs to this API: the event location address attribute (Event), the stay coordinates (Stay), and the event location attribute with the stay coordinates given as a bias (StayEvent). We also considered a more precise version of Event, which produces a result only if the geocoder returns a single unambiguous location (EventSingle). Finally, we devised the method StayEvent+Stay, which returns the StayEvent result if it exists, and the Stay result otherwise.
For each input, the geocoder gave either no result (M), a false
result (F), a true place (T), or just a true address (A). For instance,
an event occurring in “Hôtel Ritz Paris” is true if the output is for
instance “Ritz Paris”, while an output of “15 Place Vendôme, Paris”
would count as a true address. For comparison, we also evaluated
the place given by Google Timeline [
Due to limitations of the API, geocoding from stay coordinates mostly yielded address results (99.2% of the time). To better evaluate these methods, we computed the number of times in which the output was either a true place, or a true address (denoted T|A). For those methods that did not always return a result, we computed a precision metric PT|A (resp., PT), that is equal to the ratio of T|A (resp., T) to the number of times a result was returned. We computed a F1-measure based on the PT|A precision, and a recall assimilated to the number of times the geocoder returned a result (1 − M).
The evaluation was performed on 250 randomly picked (stay, event) pairs in Barack’s dataset. The results are shown in Table 3. The Google Timeline gave the right place or address only 17.2% of the time. The EventSingle method, likewise, performs poorly, indicating that the places are indeed highly ambiguous. The best precision (PT|A of 67.2%) is obtained by Geocoding with the StayEvent, but this method returns a result only 50.0% of the time. StayEvent+Stay, in contrast, can find the right place 28.4% of the time, and the right place or address 50.0% of the time, which is our best result. We are happy with this performance, considering that around 45% of the event locations were room numbers without mention of a building or place name (i.e., C101). 5.4
The user can query the KB using SPARQL [
Hub of personal knowledge. Finally, the user can use the bidirectionality of synchronization to enrich her existing services. For instance, she can enrich her personal address book (CardDav) with knowledge inferred by the system (e.g., a friend’s birth date extracted from Facebook) using a SPARQL/Update query. 6
We now review the related work, on personal information management, on information integration, and on the specific tasks of location analysis and calendar matching. 6.1
This work is motivated by concept of personal information
management (PIM), taking the viewpoint of [
Finding and Re-finding. PIM has been concerned with improving how individuals go about retrieving a piece of information to meet a particular need.
For searching within the user’s personal computer, desktop
fulltext search tools capable have been developed for various operating
systems and platforms [
Other research eforts have focused on ameliorating the process
of finding things the user has already seen, using whatever context
or meta-information that the user remembers [
Keeping. PIM has also addressed the question: What kind of information should be captured and stored in digital form?
A central idea of [
Diferent from lifelogging, which does not focus on the analysis
of the logged information, the quantified self is a movement to
incorporate data acquisition technology on certain focused aspects
of one’s daily life [
Organizing. PIM also deals with the management and organization of information. It is for instance concerned with the management of privacy, security, distribution, and enrichment of information.
A personal data service (PDS) lets the user store, manage and
deploy her information in a structured way. The PDS may be used
to manage diferent identities and/or as central point of information
exchange between services. For instance, an application that
recommends new tracks based on what the user likes to listen may need
to use adapters and authenticate with diferent services keeping
a listening history. Instead, the PDS centralizes this information
and the application only needs an adapter to connect to this PDS.
Higgins [
MyData [
Organizing information as a time-ordered stream of documents
(a lifestream) has been proposed as a simple scheme for reducing
the time the user spends in manually organizing documents into
a classic hierarchical file system [
For managing personal information, diferent levels of
organization and abstraction have been proposed. Personal data lakes [
Some proprietary service providers, such as Google and Apple,
have arguably come quite close to our vision of a personal
knowledge base. They integrate calendars, emails, and address books,
and allow smart exchanges between them. Some of them even
provide intelligent personal assistants that proactively interact with the
user. However, these are closed source proprietary solutions that
promote vendor lock-in. In response, open-source alternative
solutions have been developed, to cloud storage in particular, such as
ownCloud [
Data matching (also known as record or data linkage, entity
resolution, object/field matching) is the task of finding records that refer
to the same entity across diferent sources. It is extensively utilized
in data mining projects and in large-scale information systems by
business, public bodies and governments [
Common standards, such as vCards and iCalendar, have
advanced the state of the art by allowing provider-independent
administration of personal information. There is also a proposed standard
for mapping vCard content and iCalendars into RDF [
The ubiquity of networked mobile devices able to track users’
locations over time has been greatly utilized for estimating trafic
and studying mobility patterns in urban areas. Improvements in
accuracy and battery eficiency of mobile location technologies
have made possible the estimation of user activities and visited
places on a daily basis [
The RDF model seems well-suited for building a personal knowledge base. However powerful, making full use of it relies on being able to write and run performant queries. At this point, we cannot focus on optimizing for a specific application. It is not yet clear up to what extent Thymeflow should hold raw data (such as the entire location history), which, depending on how it is used, may be loaded on demand, in mediation style. Additionally, we would like to raise the following issues that could drive future research and development: • Gathering data for experiments: The research community might benefit from building and maintaining sets of annotated multi-dimensional personal information for use in different kinds of tasks. This is challenging, specially due to privacy concerns. • Opportunities: Internet companies that already hold a lot of user data are not yet integrating everything they have in a coherent whole, and are not performing as well as we think they could. For instance, Google Location History does not integrate the user’s calendar, unlike we do. We think that there are still many opportunities to create new products and functionalities from existing data alone. • Inaccessible information: The hard truth is that many popular Internet-based services still do not provide an API for conveniently retrieving user data out of them, or that such an API is not feature-complete (e.g., Facebook, WhatsApp). 8
The Thymeflow system integrates data from emails, calendars, address books, and location history, providing novel functionalities on top of them. It can merge diferent facets of the same agent, determine prolonged stays in the location history, and align them with events in the calendar.
Our work is a unique attempt at building a personal knowledge base. First, the system is complementary to and does not pretend to replace the existing user experience, applications, and functionalities, e.g., for reading/writing emails, managing a calendar, organizing files. Second, we embrace personal information as being fundamentally distributed and heterogeneous and focus on the need of providing knowledge integration on top for creating completely new services (query answering, analytics). Finally, while the system could benefit from more advanced analysis such the extraction of entities from rich text (e.g., emails), for linking them with elements of the KB, our first focus is on enriching existing semi-structured data, which improves the quality of data for use by other services.
Our system can be extended in a number of directions, including incorporating more data sources, extracting semantics from text, complex analysis of users’ data and behavior. Future applications include personal analytics, cross-vendor search, intelligent event planning, recommendation, and prediction. Also, our system could use simpler query language, perhaps natural language, or even proactively interact with the user, in the style of Apple’s Siri, Google’s Google Now, Microsoft’s Cortana, or Amazon Echo.
While the data obtained by Thymeflow remains under the user’s direct control, fully respecting her privacy, the data residing outside of it may not. However, using Thymeflow, the user could have a better understanding of what other systems know about her, which is important first step in gaining control about it.