Convenience-Based Periodic Composition of IoT Services

Abstract

We propose a novel service mining framework to personalize services in an IoT based smart home. We describe a new technique based on the concept of convenience to discover periodic composite IoT services to suit the smart home occupant’s convenience needs. The key features of convenience is the ability to model the spatio-temporal aspects as occupants move in time and space within the smart home. We propose a novel framework for the transient composition of spatio-temporal IoT service. We design two strategies to prune non-promising compositions. Specifically, a significance model is proposed to prune insignificant composite IoT services. We describe a spatio-temporal proximity technique to prune loosely correlated composite IoT services. A periodic composite IoT service model is proposed to model the regularity of composite IoT services occurring at a certain location in a given time interval. The experimental results on real datasets show the efficiency and effectiveness of our proposed approach.

1 Introduction

The Internet is evolving from interconnecting computers to interconnecting things [1]. The Internet of Things (IoT) paradigm enables physical devices to connect and exchange information. IoT devices allow objects to be sensed or controlled remotely through the Internet [1]. The key challenge is that IoT devices are highly heterogeneous in terms of supporting infrastructure ranging from networking to programming abstraction [7]. Service-oriented Computing (SOC) is a promising solution for abstracting things on the Internet as services by hiding the complex and diverse supporting infrastructure [8]. This abstraction can shift the focus from dealing with technical details to how services are to be used [6]. We refer services for Internet of Things as IoT services. Daily life things such as a light is connected to the Internet and is represented as a light service.

An application domain for IoT is the smart home. A smart home can be considered as any regular home which has been augmented with various types of IoT services [15]. The purpose of a smart home is to make residents’ life more convenient and efficient [17]. Current research mainly focus on basic capabilities such as communication, computing, sensing and so on, which are indeed fundamental research topics [3]. However, these basic capabilities are not enough for IoT services. They should have more advanced intelligence i.e., the capability of understanding the physical world. To empower such high-level intelligence in smart homes, a key task is to discover periodic composition of IoT services which can represent periodic human activities. Periodic composite IoT services can be loosely defined as the composite IoT services’ repeating occurrence at certain locations with regular time intervals. For example, a resident may have the habit of taking shower around 10 pm. It is of paramount importance to discover periodic composite IoT services. Periodic composition of IoT services can provide an insightful and concise explanation of IoT service usage patterns. These patterns can be used to design intelligent control of IoT services in smart homes to reduce residents’ interactions with IoT services. Reducing those interactions provides more convenience for residents. Such periodic composite IoT services are also useful for human activity prediction. If an IoT service usage fails to follow its regular periodic composition, it could be a signal of abnormality.

Fig. 1.

Examples of periodic composite IoT services

It is challenging to provide convenience by discovering periodic composite IoT services from IoT service usage history (i.e., IoT service event sequence). For example, Fig. 1 shows the IoT service usage history (on the left). We can see that it is difficult to extract periodic composite IoT services (on the right). We identify three key challenges.

• The set of IoT services are not known, which may be used collectively to fulfill a daily task. These IoT services are spatio-temporally correlated. We refer such set of IoT services as composite IoT services.

• There are many opportunities of establishing spatio-temporal relationships among IoT services, leading to an explosive number of possible composite IoT services. Many of these composite IoT services may be insignificant and loosely correlated. As a result, there is a need to prune insignificant and loosely correlated composite IoT services.

• The associated time interval and location for the periodic composite IoT services are not known. The composite IoT service may not occur exactly at the same time in a particular location. Therefore, there is a need to estimate the associated time interval and location.

In this paper, we focus on providing convenience by discovering periodic composite IoT services. At the first stage, we focus on discovering composite IoT services. Then, we employ significance and proximity strategies to prune insignificant and loosely correlated composite IoT services. At the third stage, we estimate associated time interval and location for the candidates generated in the second stage. Lastly, we measure how much convenience can be obtained by applying discovered periodic composite IoT services. The key contributions are as follows: (1) A new IoT service model and a composite IoT service model are proposed based on spatio-temporal features. (2) A significance model is proposed to prune insignificant composite IoT services. We also propose a proximity model in terms of spatial-proximity and temporal-proximity to filter out loosely correlated composite IoT services, (3) A periodic composite IoT service model is proposed to represent the regularity of composite IoT services occurring at a certain location in a time interval, (4) A convenience model is proposed to measure the benefits of applying periodic composite IoT services, and (5) A novel algorithm PCMiner (i.e., Periodic Composite IoT service Miner) is designed to discover periodic composite IoT services from event sequences.

The rest of the paper is organized as follows. Section 2 formally defines key concepts. Section 3 details the proposed algorithm PCMiner. Section 4 shows the experimental results. Section 5 surveys the related work. Section 6 conclude the paper and highlights some future work.

Fig. 2.

(a) An example of a composite IoT service; (b) Time interval relations

Motivating Scenario

We use the smart home as our motivating scenario. Sarah lives alone in a smart home. We assume everything such as lights, TV, oven, window, and floors are connected to the Internet and represented as IoT services. This smart home aims to improve Sarah’s life convenience. Intuitively, convenience can be interpreted as a smart home system which is aware of a resident’s potential needs by understanding physical environment or situations and respond proactively at the right time and in the right place. This reduces a resident’s interactions with IoT services. Let us imagine an interesting convenient life scenario in Sarah’s home. In a weekday morning, the clock wakes Sarah up at 8 am. Then the lamp is turned on automatically. Sarah gets up and prepares stuff for taking a shower. Meanwhile, the heater in the bathroom starts to heat. When Sarah steps into the bathroom, it is already warm. While Sarah is taking shower, the music player is playing her favorite music. At the same time, the kettle in the kitchen starts to work and the coffee maker starts to make a cup of Mocha coffee. After Sarah finishes showering and grooming, she goes to the kitchen to prepare for breakfast. When she is enjoying the breakfast and coffee, the TV is turned on automatically and displays her favorite sport news. After finishing the breakfast, Sarah goes to work. The TV and all the lights are turned off automatically.

A fundamental task to provide convenience is to augment IoT services with capabilities of understanding the periodic usage of composite IoT services. Let us consider Sarah performs daily activities by interacting with IoT services. These interactions are recorded as IoT service event sequences shown in Fig. 3. For example, ${<}E^{+}E^{-},(60,75){>}$ denotes the music player is playing music from time 60 to time 75 and $E^{+}$ (resp.$E^{-}$) denotes a turn on (resp. turn off) the music player event. There exist spatio-temporal relationships among IoT services. For example, in Day 1, the relationships between the music player service (i.e., E) and the shower service (i.e., F) shows that Sarah listens to music while taking a shower. In this regard, a collection of spatio-temporally correlated IoT services may represent an activity. We refer to such set of IoT services as composite IoT services. An example of a composite IoT service is shown in Fig. 2(a).

There are many ways of establishing spatio-temporal relationships among IoT services. According to Allen’s temporal logic in Fig. 2(b), there are 57 ways of generating composite IoT services in Day 1 by a brute force approach (i.e., $C_6^{2}+C_6^{3}+C_6^{4}+C_6^{5}+C_6^{6} = 57$). Many of them may be insignificant and loosely correlated. Therefore, we design the significance and proximity strategies to filter out these insignificant and loosely correlated composite IoT services.

The proximate and significant composite IoT services can represent the resident’s daily activities. The residents usually performs his/her daily activities periodically in terms of time and location. For example, the resident usually goes to bed during 11pm to 12pm and wakes up during 8am to 9am. We refer such repeating composite IoT services at certain location with regular time intervals as periodic composite IoT services. The periodic composite IoT services can be serve as a knowledge basis for providing convenience. A convenience model is introduced to quantify how much benefits of applying periodic composite IoT services. In this paper, we focus on providing convenience by discovering periodic composite IoT services.

Fig. 3.

An example of service event sequences

2 System Model

We first introduce the notion of IoT services and composite IoT services based on spatio-temporal features [25]. Then, a significance model and a proximity model are proposed to prune non-promising composite IoT services. We introduce the notion of periodic composite IoT service to model the occurrence regularity of composite IoT services. Lastly, we provide a convenience model to quantify how much benefits of applying periodic composite IoT services.

2.1 IoT Service Model

Definition 1:

IoT Service. An IoT service $S_i$ is a tuple $S_i = {<} s_i, F_i, IS, FS {>}$, where:

• $s_i$ is a unique service identifier.

• $F_i$ is a set of functions that are offered by $S_i$.

• IS (Initial State) is a tuple ${<} s_i^{+}, st_i, sl_i{>}$, where

  • $s_i^{+}$ is a symbol of IS.

  • $st_i$ is a start-time of $S_i$.

  • $sl_i = {<}x_s,y_s{>}$ is a start location of $S_i$, where ${<}x_s,y_s {>}$ is a GPS point.

• FS (Final State) is a tuple ${<} s_i^{-}, et_i, el_i{>}$, where

  • $s_i^{-}$ is a symbol of FS.

  • $et_i$ is an end-time of $S_i$.

  • $el_i = {<}x_e,y_e {>}$ is an end location of $S_i$, where ${<}x_e,y_e {>}$ is a GPS point.

We focus on the spatio-temporal features in the remainder of this paper. Thus, the representation of an IoT service $S_i$ is simplified as ${<} (s_i^+,st_i,sl_i ), (s_i^-, et_i,el_i){>}$. For example, a light service is represented as ${<}(light^+, 7pm, (1,2) ), (light^-, 9pm, (1,2) ){>}$ which is described as lighting from 7pm to 9pm in the bedroom where (1, 2) is the GPS point in the bedroom and 7pm (resp. 9pm) is the start time (resp. end time).

2.2 Composite IoT Service Model

One IoT service may not accomplish a daily activity. Multiple IoT services may be composed to fulfill an activity [13]. These IoT services may be used collectively based on time and location correlations to accomplish a certain daily activity. We refer such spatio-temporally correlated IoT services as composite IoT services.

Definition 2:

Composite IoT Service. A composite IoT service CS is a collection of IoT services that occur frequently in a particular spatio-temporal relationships. A composite IoT service is denoted by a tuple $CS={<} S, sup(S) {>}$ where

• $S=\{{<} (s_1^+,st_1,sl_1), (s_1^-,et_1,el_1){>},..., {<} (s_n^+,st_n,sl_n), (s_n^-,et_n,el_n){>}\}$ represents n component IoT services where ${<} (s_i^+,st_i,sl_i), (s_i^-,et_i,el_i){>}$ is a component IoT service $S_i$ as defined in Definition 1 and $st_i \le st_{i+1}$ and $st_i \le et_i$. An example of a composite IoT service is shown in Fig. 2(a). By ordering all elements $s_i^*$ (* can be + or −) in S in a non-decreasing order based on its associated time information $st_i$(or $et_i$), we can transform S into the following representation $S = {<}Seq,T, L{>}=\left\{ \begin{aligned} \alpha _1&...&\alpha _i&...&\alpha _{2n}\\ t_1&...&t_i&...&t_{2n} \\ l_1&...&l_i&...&l_{2n} \end{aligned} \right\} ,$ where $Seq = \{\alpha _1...\alpha _i...\alpha _{2n}\}$ is a symbol sequence and $\alpha _i = s_j^*$ (* can be + or −), $T=\{ t_1...t_i...t_{2n} \}$ is the time information and $t_i\le t_{i+1}$, and $L=\{l_1...l_i...l_{2n}\}$ is the location information. For example, the composite IoT service in Fig. 2(a) can be represented as $\left\{ \begin{aligned} s_2^{+ }&s_1^{+}&s_3^{+}&s_1^{-}&s_3^{-}&s_2^{-} \\ 48&50&58&65&70&75 \\ l_2&l_1&l_3&l_1&l_3&l_2 \end{aligned} \right\}$ (i.e., $l_1=(1,2 )$, $l_2=(2,4 )$, $l_3=(3,5 )$)

• sup(S) is the support for S. The support sup is the total number of occurrence in a database $S = {<}Seq, T, L{>}$ and $S' = {<}Seq', T', L'{>}$ be two composite IoT services. S is referred to as a sub-composite IoT service of $S'$ , denoted as $S \sqsubseteq S'$, if $Seq = \{ \alpha _1...\alpha _i...\alpha _{2n}\}$ is a subsequence of $Seq'=\{ \alpha _1'...\alpha _i'...\alpha _{2m}'\}$, denoted as $Seq \sqsubseteq Seq'$ with $n\le m$. $Seq \sqsubseteq Seq'$ is satisfied if there exist integers $1 \le k_1 \le k_2...k_n\le k_{2m}$ such that $\alpha _1 \subseteq \alpha _{k_1}'$, $\alpha _2 \subseteq \alpha _{k_2}', \ldots , \alpha _n \subseteq \alpha _{k_n}'$. Given an IoT service event sequence DB, the tuple $(sid, S')$ (i.e., sid is a sequence ID and $S'$ is the composite IoT service) is said to contain a sub-composite IoT service S if $S \sqsubseteq S'$. The support sup of S in DB, denoted as sup(S) is the number of tuples containing S. sup(S) can be formalized as follows.

  $$\begin{aligned} sup(S) = | \{ (sid, S)\in DB | S\sqsubseteq S' \}| \end{aligned}$$

  (1)

2.3 Significance and Proximity Model for Composite IoT Services

There are many possibilities of establishing spatio-temporal relationships among IoT services, leading to an explosive generation of composite IoT services. Many of the composite IoT services are insignificant and loosely correlated. Thus, there is a need to filter out these non-promising composite IoT services. We explore this problem from two aspects. On the one hand, those composite IoT services that occur frequently are more likely to be significant that those occur less frequently. Thus, we propose a significance model to quantify how much significance these composite IoT services are from the statistic aspect. On the other hand, the IoT services that occur proximately in terms of time and location are more likely to be correlated. For example, from the spatial perspective, the relationship between the TV and the light in the same dining room may reveal a high correlation between these two IoT services. However, the co-occurrence of using TV in the dining room and using the light in the bedroom may merely be a coincidence. From the temporal perspective, using the TV and the light in the evening reveal that there may exist a high correlation between the two IoT services during that time. However, using the TV in the evening and using the light in the morning may not have any correlation. In this regard, we use proximity to characterizes correlation strength among component IoT services in terms of spatial-proximity and temporal-proximity. By spatial-proximity, it characterizes the location correlation strength among component IoT services. By temporal-proximity, it characterizes the temporal correlation strength among component IoT services. The proximity model is adapted from the approach for measuring spatio-temporal interval data distance [11].

Definition 3:

Significance. Significance is used for evaluating statistic importance of CS. Given a composite IoT service $CS = {<}S, sup{>}$, its significance is formalized as:

$$\begin{aligned} significance (S) = \frac{ \sqrt{expect(S)}}{sup(S)-expect(S)} \end{aligned}$$

(2)

where expect(S) is the expected number of occurrence in a DB. To estimate expect(S), we adapt the statics model proposed in [14] by considering IoT services’ various usage frequency across different regions in smart homes [16]. In practice, IoT services’ usage frequency may vary across different regions in smart homes. For example, if a resident spends most of his/her time in their living room during the day and only goes to their bedroom for sleeping, then IoT services in the living room will be used more frequently than those in the bedroom. Thus, composite IoT services for sleeping may be ignored when searching for frequent composite IoT services.

Given a composite IoT service CS whose symbol sequence is $Seq = \{ s_1^*...s_i^*...s_m^*\}$, suppose that Seq is a possible outcome drawn from the symbol set $A=\{ a_1...a_i...a_{n}\}$ with $P(a_i)$ following Bernoulli distribution such that $\sum _{i=1}^{n}P(a_i)=1$. Given a DB and a region set $R=\{r_1,...,r_k\}$, $DB_{r_i}$ records IoT service event sequences occurring at region $r_i$. $Num(a_i)_{DB_{r_i}}$ is the number of the event $a_i$ occurrence in database $DB_{r_i}$. $P(a_i)$ can be estimated by $\frac{Num(a_i)_{DB_{r_i}}}{\sum _{a_j \in A }{Num(a_j)_{DB_{r_i}} }}$. Therefore, the occurrence probability P(S) and expect(S) can be formalized as follows.

$$\begin{aligned} P(S)=\prod _{ \forall s_i^{+} \in Seq} P(s_i^{*}) \end{aligned}$$

(3)

$$\begin{aligned} expect(S)=P(S)\cdot |DB_{r_i}| \end{aligned}$$

(4)

where $|DB_{r_i}|$ is number of IoT service events in region $r_i$. Note that we only count $P(s_i^{*})$ once because each $s_i^{*}$ has two points (i.e., $s_i^{+}$ and $s_i^{-}$).

Definition 4:

Proximity. Given a composite IoT service $CS = {<} S, sup{>}$ where S is in the form $S = {<}Seq,T, L{>}=\left\{ \begin{aligned} \alpha _1&...&\alpha _i&...&\alpha _{2n}\\ t_1&...&t_i&...&t_{2n} \\ l_1&...&l_i&...&l_{2n} \end{aligned} \right\} ,$ its proximity function is defined as follows.

$$\begin{aligned} U = w_1\cdot spatial\_proximity + w_2\cdot temporal\_proximity \end{aligned}$$

(5)

where $w_i(i=1, 2)$ is a weight such that $w_i \in [0,1]$ and $w_1 + w_2 =1$. The spatial_proximity and temporal_proximity are formalized as follows.

• Spatial_proximity: The spatial_proximity measures the average location proximity of all composite IoT service instances. The $spatial\_proximity$ for a composite IoT service instance is first formalized in Eq. (6). Then the average $spatial\_proximity$ for the composite IoT service is formalized in Eq. (7).

  $$\begin{aligned} Spa = \sum _{i=1}^{n} \frac{1}{|x_i-x_{i+1}|+|y_i-y_{i+1}| } \end{aligned}$$

  (6)
  $$\begin{aligned} spatial\_proximity = \frac{\sum _{j=1}^{sup} Spa_j}{sup} \end{aligned}$$

  (7)

  where n is the total number of component IoT services, $l_i={<}x_i, y_i {>}$ and $l_{i+1} = {<}x_{i+1}, y_{i+1} {>}$ are two locations for two consecutive component IoT services, and sup is the support for the composite IoT service. For example, for the composite IoT service in Fig. 2(a), its spatial_proximity score is $Spa= \frac{1}{|l_2-l_1|} + \frac{1}{|l_1-l_3|} = \frac{1}{|2-1|+|4-2| } + \frac{1}{|1-3|+|2-5| } =0.53$. In this paper, we use Manhattan distance proposed in [12] to measure the proximity because it is computing efficient.

• Temporal_proximity: The temporal_proximity measures the average temporal proximity of all composite IoT service instances. We adapt the technique of evaluating the distance between time-interval based data in [11]. For each component IoT service instance $S_i = \{ (s_i^+, st_i, sl_i), (s_i^-, et_i,el_i)\}$, we utilize a function $f_i$ with respect to t to map the temporal aspect of $S_i$. $f_i$ is formalized as follows.

  $$\begin{aligned} f_i(t)= \left\{ \begin{aligned} 1&,&t \in [st_i, et_i] \\ 0&,&otherwise \end{aligned} \right. \end{aligned}$$

  (8)

  Then we have a set of functions $\{ f_1, f_2,...f_n\}$ corresponding to the composite IoT service instance. The temporal_proximity for the composite IoT service instance is calculated by Eq. (9). The average temporal_proximity is calculated by Eq. (10).

  $$\begin{aligned} Temp = \frac{\int _{t_1}^{t_{2n}} \sum _{i=1}^nf_i(t)\,dt }{(t_{2n}-t_1)\cdot n} \end{aligned}$$

  (9)
  $$\begin{aligned} temporal\_proximity = \frac{\sum _{j=1}^{sup} Temp_j}{sup} \end{aligned}$$

  (10)

  where $t_1$ and $t_{2n}$ are the first and the last time information of CS, respectively, and n is the number of component IoT services. For example, the temporal_proximity score of the composite IoT service {${<}$stove, [18:00, 19:00]>, <washing machine, [18:40, 19:20 ]>} can be calculated as $\frac{(18:40-18:00)+ (19:00-18:40)\cdot 2+(19:20-19:00)}{(19:20-18:00)\cdot 2}$ = 0.625. This composite IoT service can be interpreted as when the resident is cooking, he/she is also doing laundry. Another composite IoT service is {<stove, [18:00, 19:00]>, <fan, [18:00, 19:00]>} and its temporal_proximity score is 1. Thus the latter composite IoT service is considered to be more temporally proximate than the former.

2.4 Periodic Composite IoT Service Model

In this section, we introduce the novel notion of periodic composite IoT service to model the regularity of repeating composite IoT services.

Definition 5:

Periodic composite IoT service. A periodic composite IoT service PC is defined as the repeating composite IoT services at certain locations with regular time intervals. It is denoted by a tuple $PC = {<} CS, T, L, P{>}$ where

• CS is a composite IoT service.

• $T = {<}T_s, T_e{>}$ is a representative time interval associated with CS, where $T_s$ and $T_e$ are the start time and end time of CS, respectively. Suppose all start time and end time of CS in DB constitutes the set $\tau = \{ {<}st_1, et_1{>}, {<}st_2,et_2{>}... {<}st_m,et_m{>}\}$. We need to find the representative time interval ${<}T_s, T_e{>}$ which minimizes the dissimilarity between the instance ${<}st_i, et_i{>}$. We define the dissimilarity dis between two time intervals.

  $$\begin{aligned} dis=|T_s-st_i|+|T_e-et_i| \end{aligned}$$

  (11)

  Thus, the total dissimilarity between ${<}T_s, T_e{>}$ and $\tau$ can be defined by Eq. (12).

  $$\begin{aligned} Dis(T, \tau )= \sum _{i=1}^{m} |T_s-st_i|+|T_e-et_i| \end{aligned}$$

  (12)

  To minimize $Dis(T, \tau )$, Eq. (12) can be transformed into two known minimization problems, that is, find $T_s$ and $T_e$ to minimize $\sum _{i=1}^{m} |T_s-st_i|$ and $\sum _{i=1}^{m} |T_e-et_i|$, respectively. $T_s$ is the median of the start time set $\{ st_1, st_2... st_m\}$ and $T_e$ is the median of the end time set $\{ et_1, et_2... et_m\}$. The proof can be found in [19].

• L is the region location of CS such as the bedroom and the bathroom.

• P is the probability of CS occurring around time interval T at location L. Suppose the time information of a CS instance is ${<}st_j, et_j{>}$. The CS is said to occur around time interval T if their dissimilarity dis is no more than a tolerance threshold $\zeta$, that is, $|T_s-st_j|+|T_e-et_j| \le \zeta$. P can be formalized as follows.

  $$\begin{aligned} P=\frac{Num}{TNum} \end{aligned}$$

  (13)

where Num is the number of CS occurrence around time interval T at location L. TNum is the total number of CS occurrence in the database.

2.5 Convenience Model

The discovered periodic composite IoT services can be served as knowledge basis for building an intelligent system to provide convenience for the residents. By convenience, it is interpreted as the benefits of applying periodic composite IoT services via reducing residents’ interactions with IoT services. The convenience can be quantified as follows.

Definition 6:

Convenience. Given an IoT service event sequence $\{{<}(a_1^+, st_1, sl_1), (a_1^-, et_1, el_1) {>}\ldots {<} (a_n^+, st_n, sl_n), (a_n^-, st_n, sl_n){>}\}$ during the time period [$st_1, et_n$], suppose this sequence is initialized from a set of periodic composite services { $PC_1, PC_2 \ldots PC_m$}. According to the representative time information of the $PC_m$, we can roughly estimate the next $PC_{m+1}$ occurrence. The IoT service events involved in $PC_{m+1}$ is { $b_1$, $b_2$...$b_m$ }. Suppose the actual event set occurs next is { $c_1, c_2 \ldots c_k$ }. Therefore, the amount of convenience can be quantified by Eq. (14).

$$\begin{aligned} convenience = \frac{|\{b_1, b_2...b_m\}\cap \{c_1, c_2...c_k\}| }{|\{ c_1, c_2...c_k\}|} \end{aligned}$$

(14)

where $|\{c_1, c_2...c_k\}|$ is the number of events and $|\{b_1, b_2...b_m\}\cap \{c_1, c_2...c_k\}|$ is the number of correctly estimated events.

3 Discovering Periodic Composite IoT Service Approach

We develop the algorithm PCMiner to efficiently discover periodic composite IoT services from IoT service event sequences. Algorithm 1 shows the details of PCMiner. The algorithm consists of four phases. The mining process starts with dividing the search space. Then PCMiner searches all composite IoT services in a determined space. Third, PCMiner applies the significance and proximity strategies to remove non-promising composite IoT services. Finally, PCMiner collects time information and location information for candidates generated in the third phase. Based on these information, time period and location corresponding to the candidates are estimated which leads to generating a set of periodic composite IoT services. For the sake of consistency with the terms from data analysis techniques, we use event patterns to refer to composite IoT services occurring in service event sequences. We use the running example in Fig. 3 to illustrate the process of PCMiner shown in Fig. 4.

Fig. 4.

The Process of PCMiner

Phase I: Dividing Search Space. The layout of a smart home consists of multiple regions such as a bedroom and a kitchen. Each IoT service event is associated with a region. For example, turn on the lamp event occurs in the bedroom. Given a region set $r= \{ r_1,r_2...r_n\}$, we divide the database DB into multiple smaller databases $DB_{r_i}$. Each $DB_{r_i}$ records IoT service event sequences occurring in the region $r_i$. In later phases, the discovering process performs on each sub-databases. For the purpose of illustrating PCMiner, we assume all IoT service event sequences in the running example shown in Fig. 3 are from the same region and constitute a sub-database.

Phase II: Searching Event Patterns. PCMiner employs a divide-and-conquer, pattern-growth principle from Prefixspan [23] as follows: event sequence databases are recursively projected into a set of smaller projected databases based on the current event patterns. Event patterns are then grown by searching the smaller projected databases.

Definition 7:

Projected database. Let p be an event pattern in a database DB. The p-projected database, denoted as $DB|_p$, is the collection of suffixes of event sequences in DB with regard to the prefix p.

The searching process consists of three sub-phases. PCMiner first finds the set of 1-length event patterns. Then, PCMiner constructs projected databases for each 1-length event pattern generated in the phase one. Third, the event patterns are grown by searching their corresponding projected databases. Each of these sub-phases is detailed as follows.

1. Find the set of 1-length event patterns $L_1$. Given a database as shown in Fig. 3, PCMiner first scans the database to count the number of each event pairs and discards those events whose support is less than the minimum support threshold. If the minsup threshold is 2, all discovered 1-length event patterns whose support is not less than 2 constitute the 1-length event pattern set $L_1$. For example, in Fig. 4, ($A^+A^-$): 3 denotes the event pattern and its associated support count.

2. Construct projected databases for each 1-length event pattern. Let $L_1 = \{ \alpha _1^1, \alpha _2^1 ... \alpha _n^1\}$ be the set of 1-length event patterns. For each $\alpha _i^1$, a corresponding projected database $DB|_{\alpha _i^1}$ is created. $DB|_{\alpha _i^1}$ is a collection of suffix event sequences with regard to the prefix $\alpha _i^1$.

3. k-length event pattern $\alpha$ is grown to the (k + 1)-length event pattern $\alpha '$ through searching the projected database $DB|_{\alpha }$ corresponding to $\alpha$ ($k \ge 1$). For a prefix $\alpha$, PCMiner scans its projected database $DB|_{\alpha }$ once to find the set of local frequent event pairs {$e_1, e_2 \ldots e_n$} and discards infrequent ones. Note that since event pairs are counted, these single events in the projected database will not be counted again. Frequent event pairs $e_i$ are appended to the prefix $\alpha$, generating the new frequent event pattern $\alpha '$ with the length increased by 1. Therefore, the set of (k + 1)-length event patterns prefixed with $\alpha$ are generated.

We illustrate the process of finding event patterns prefixed with $(A^+A^-)$. By scanning the $(A^+A^-)$ projected database $DB|_{(A^+A^-)}$, its local frequent event pairs are ($B^+B^-$: 3), ($C^+C^-$: 3), ($E^+E^-$: 2), and ($F^+F^-$: 3). Thus, the set of all 2-length event patterns $L_2$ prefixed with $(A^+A^-)$ are found, and they are: ($A^+A^-B^+B^-$: 3), ($A^+A^-C^+C^-$: 3), ($A^+A^-E^+E^-$: 2), and ($A^+A^-F^+F^-$: 3). Recursively, all 2-length event patterns are used to find 3-length event patterns by constructing and searching their projected databases. By projecting ($A^+A^-B^+B^-$), we find frequent event pairs from its projected database which are ($C^+C^-$: 3), ($E^+E^-$: 2), and ($F^+F^-$: 3). By appending these frequent event pairs to the prefix ($A^+A^-B^+B^-$), we have 3-length event patterns ($A^+A^-B^+C^+C^-B^-$: 3), ($A^+A^-B^+B^-E^+E^-$: 2), and ($A^+A^-B^+B^-F^+F^-$: 3). Similarly, we find ($A^+A^-C^+C^-E^+E^-$: 2) and ($A^+A^-C^+C^-F^+F^-$: 3) prefixed with ($A^+A^-C^+C^-$), and ($A^+A^-E^+F^+F^-E^-$: 2) prefixed with ($A^+A^-E^+E^-$). We find $L_4$ and $L_5$ in the same approach.

Phase III: Calculate Significance and Proximity for Event Patterns. For each event pattern generated in phase II, we collect the time information and location information from the event sequences. Based on these information, we calculate the statistic significance for each event pattern by Definition 3. We also discard insignificant ones if its significance is less than the significance threshold minsig. Given proximity threshold minpro, we calculate average proximity for each event pattern by Definition 4 and filter out those patterns whose proximity are less than minpro. For the running example, if the weight for $spatial\_proximity$ and $temporal\_proximity$ is set to be 0 and 1, respectively, the proximity of all 2-length event patterns are $Prox( A^{+}A^{-}B^{+}B^{-})=0.418$, $Prox( A^{+}A^{-}C^{+}C^{-})=0.358$, $Prox( A^{+}A^{-}E^{+}E^{-})=0.098$, $Prox( A^{+}A^{-}F^{+}F^{-})=0.082$, $Prox( B^{+}C^{+}C^{-}B^{-})=0.906$, $Prox( B^{+}B^{-}E^{+}E^{-})=0.224$, $Prox( B^{+}B^{-}F^{+}F^{-})=0.215$, $Prox(C^{+}C^{-}E^{+}E^{-})=0.209$, $Prox( C^{+}C^{-}F^{+} F^{-})=0.199$, $Prox( E^{+}F^{+}F^{-}E^{-})=0.879$. If the minpro is set to be 0.3, event patterns whose proximity is less than 0.3 are filtered out. The ultimate outcomes of our example are ${<} E^{+}F^{+}F^{-}E^{-}{>}$ and ${<} A^{+}A^{-}B^{+}C^{+}C^{-}B^{-}{>}$ and their respective proximity are 0.879 and 0.494.

Phase IV: Generating Periodic Event Patterns. After performing phase III, we obtain significant and proximate event patterns. Based on the time and location information collected in phase III, the algorithm estimates the time period, location, and probability for each event pattern by Definition 5. The outcomes of this phase is a set of event patterns associated with time intervals, location, and a probability (i.e., periodic composite IoT services).

4 Experimental Results

We systematically evaluate the approach proposed in this paper. The language used is Java and the experiments are performed on a 1.6 GHz AMD processor and 2 GB RAM under Windows 7. We evaluate the proposed approach using three real datasets, namely, Data1, Data2, Data3. Specifically, Data1 and Data2 are from CASAS datasets, which are collected in smart home environment [18]. For location information, we refer to the layout of sensors attached on objects for grouping objects into corresponding locations. Data1 and Data2 are in the format of <date, time stamp, sensor ID, on/off> (e.g., <2008-02-27, 12:46:37, M13, OFF >). Data3 is collected from a single apartment for two weeks [10]. The Data3 are in the format of <id, start time, end time, location> (e.g., <light, 7:00, 8:00, bedroom>). In addition, all datasets are annotated with corresponding daily activities. There are 5 and 8 activities in Data1 and Data2, respectively. For Data3, 23 activities are recorded and annotated. We conduct four sets of experiments. The first set is to evaluate the performance and scalability of PCMiner. The second set is to evaluate the effectiveness of the pruning strategies (i.e., significance and proximity). The third set is to evaluate the applicability of the proposed approach by showing the discovered periodic composite IoT services from real datasets. The fourth set is to measure the convenience by applying the discovered periodic composite IoT services.

Fig. 5.

Performance and scalability of PCMiner

The first set of experiments is conducted on a dataset which is a combination of three datasets i.e., Data1, Data2, and Data3. We vary the support threshold sup from 4% to 10%. Figure 5(a) shows the execution time of PCMiner decreases by increasing the support threshold. Figure 5(b) illustrates that the number of discovered event patterns decreases by increasing the support.

In the second set of experiments, we assess the effectiveness of significance and proximity in pruning non-promising event patterns. Similar to our previous experiment, we use the combined dataset. We set the significance to be 0.01. We test the effectiveness of significance in reducing insignificant IoT service event patterns while varying different support threshold. Figure 6(a) depicts the number of discovered patterns and significant patterns at different support threshold. The results show that the significance strategy performs effectively in pruning insignificant event patterns, which is an expected results. For example, the significance strategy can prune event patterns from 2954 to 1108 at the 5% support threshold. In addition, we test the effectiveness of proximity strategy in filtering out loosely correlated event patterns. Since the GPS point of each service in not available in the datasets, we set the weight of $spatial\_proximity$ to be 0 and the weight of $temporal\_proximity$ to be 1. We set the proximity to be 0.39. Figure 6(b) illustrates the number of discovered event patterns and proximate event patterns at different support threshold. The results show that the proximity strategy is effective. For example, the proximity strategy can prune loosely correlated event patterns from 2954 to 1705 at the 5% support threshold. These are expected results because the significance and proximity strategies enable PCMiner to filter out non-promising event patterns in each iteration and the search scope is shrunk for the next iteration.

Fig. 6.

Effectiveness of significance and proximity strategies

We perform the third set of experiments on Data3 to evaluate the applicability of our proposed approach. Table 1 shows the primary discovered composite IoT services. Some of the composite IoT services are indeed difficult to be discovered because they are less frequent. For example, the “lawn work” and “going out for entertainment” compositions occur only once during two weeks. Next, we check the discovered periodic compositions. Ideally, we want to associate one time interval with a composition. However, we discover that some IoT service compositions may associate multiple time intervals. For example, the “taking medication” service composition occurs in the morning and in the evening. This is a very practical issue. In this experiment, we group the discovered composition instances using a preliminary technique, that is, two time intervals are grouped together if they overlap. The tolerance threshold $\zeta$ is set to be 2 h. We can see from Table 1 that the resident performs some activities regularly. For example, one of the striking periodic activities is the “preparing breakfast”. There is 75% chance that he/she will “prepare breakfast” during 5:16 and 6:51.

We conduct the fourth set of experiments on Data3 to measure how much convenience can be obtained by applying the discovered results in Table 1. We showcase some preliminary results. For example, based on the representative time, the “preparing breakfast” is likely to be followed by “watching TV” and “watching TV” is likely to be followed by “going out for shopping”. Given the “preparing breakfast” activity on 4/22/2003, we can obtain 50% convenience.

Table 1. Primary discovered periodic composite IoT services from Data3

5 Related Work

A Web service mining framework is proposed to discover interesting composite services from available services [4]. This framework models Web services using ontologies and an efficient algorithm is proposed to discover composite services. In [24], a graphic model is proposed to represent the dependency among services. A general service mining framework is proposed based on an ontology service model [5]. Service relationships are established via ontology attributes. A Correlation Degree method is presented to evaluate the correlation strength among services. Most of existing work consider service relationships based on their input/output correlations, pre/post condition correlations etc. However, in the context of IoT, spatio-temporal relationships among IoT services are implicit and subtle. In [12], an efficient algorithm CoPMiner is developed to mine the temporal relationships among appliances in the smart home environment. The key idea of CoPMiner is to transform interval-based event sequences into endpoint based sequences. It also reformulates the problem of discovering temporal patterns among appliances as discovering frequent patterns from endpoint sequences. Location information regarding appliances is utilized to filter out insignificant temporal patterns. However, temporal distance is not considered in [12], which may result in undesirable frequent temporal patterns. In [9], an efficient algorithm IEMiner is proposed to discover temporal patterns for classification. In [2], a novel graph-based approach is proposed to capture the subtle relationships among things based on things’ usage time, location, and users’ social network information. The Random Walk with Restart method is applied to discover things relationships.

There are many research on human activity discovery. In [16], an efficient algorithm COM is proposed to discover human activity pattens from sensor event data. These patterns are used to build a HMM model for recognizing human activities. In [20], a probabilistic and Markov chain approach is proposed to discover complex human activity patterns. These patterns associated with context information are used to recognize activities. A general framework is proposed to address the problem of complex activity prediction by mining temporal sequence patterns from video [21]. A probabilistic suffix tree model is introduced to model activities. There have been little research into the human activity recognition which considers the periodic feature. For example, [22] discovers periodic activities from trajectory data such as staying in the office during daytime and staying at home in the evening.

6 Conclusion and Future Work

We addressed the problem of discovering periodic composite IoT services to provide personalized convenience to residents. An IoT service model and a composite IoT service model are proposed in terms of spatio-temporal aspects. The experimental results show our proposed significance and proximity strategies are effective in pruning non-promising composite IoT services. The periodic composite IoT service model is introduced and is applied to provide convenience. We introduce a new algorithm PCMiner to discover periodic composite IoT services. Future work includes improving the performance of PCMiner. Furthermore, we will apply the discovered periodic composite IoT services to build an intelligent system for providing convenience.