How to Use TemporalDriven Constrained Clustering to Detect Typical Evolutions
Abstract
In this paper, we propose a new timeaware dissimilarity measure that takes into account the temporal dimension. Observations that are close in the description space, but distant in time are considered as dissimilar. We also propose a method to enforce the segmentation contiguity, by introducing, in the objective function, a penalty term inspired from the Normal Distribution Function. We combine the two propositions into a novel timedriven constrained clustering algorithm, called TDCKMeans, which creates a partition of coherent clusters, both in the multidimensional space and in the temporal space. This algorithm uses soft semisupervised constraints, to encourage adjacent observations belonging to the same entity to be assigned to the same cluster. We apply our algorithm to a Political Studies dataset in order to detect typical evolution phases. We adapt the Shannon entropy in order to measure the entity contiguity, and we show that our proposition consistently improves temporal cohesion of clusters, without any significant loss in the multidimensional variance.
Received (Day Month Year)
Revised (Day Month Year)
Accepted (Day Month Year)
Keywords: semisupervised clustering; temporal clustering; temporalaware dissimilarity measure; contiguity penalty function; temporal cluster graph structure.
1 Introduction
Researchers in Social Sciences and Humanities (like Political Studies) have always gathered data and compiled databases of knowledge. This information often has a temporal component, the evolution of a certain number of entities is recorded over a period of time. Each entity is described using multiple attributes, which form the multidimensional description space. Therefore, an entry in such a database would be an observation, a triple . An observation signifies that the entity is described by the vector at the moment of time . We denote by the entity to which the observation is associated. Similarly, is the timestamp associated with the observation . Each observation belongs to a single entity and, consequently, each entity is associated with multiple observations, for different moments of time. Formally:
For example, a database studying the evolution of democratic states will store, for each country and each year, the value of multiple economical, social, political and financial indicators. The countries are the entities, and the years are the timestamps.
Starting from such a database, one of the interests of Political Studies researchers is to detect typical evolution patterns. There is a double interest: a) obtaining a broader understanding of the phases that the entity collection went through over time (e.g. detecting the periods of global political instability, of economic crisis, of wealthiness etc.); b) constructing the trajectory of an entity through the different phases (e.g. a country may have gone through a period of military dictatorship, followed by a period of wealthy democracy). The criteria describing each phase are not known beforehand (which indicators announce a world economic crisis?) and may differ from one phase to another.
We address these issues by proposing a novel temporaldriven constrained clustering algorithm. The proposed algorithm partitions the observations into clusters , that are coherent both in the multidimensional description space and in the temporal space. We consider that the obtained clusters can be used to represent the typical phases of the evolution of the entities through time. Figure 1 shows the desired result of our clustering algorithm. Each of the three depicted entities ( and ) is described at 10 moments of time (). The 30 observations of the dataset are partitioned into 5 clusters (). In Figure (a)a we observe how clusters are organized in time. Each of the clusters has a limited extent in time, and the time extents of clusters can overlap. The temporal extent of a cluster is the minimal interval of time that contains all the timestamps of the observations in that cluster. The entities navigate through clusters. When an observation belonging to an entity is assigned to cluster and the anterior observation of the same entity is assigned in cluster , then we consider that the entity has a transition from phase to phase . Figure (b)b shows how the series of observations belonging to each entity are assigned to clusters, thus forming continuous segments. This succession of segments is interpreted as the succession of phases through which the entity passes. For this succession to be meaningful, each entity should be assigned to a rather limited number of continuous segments. Passing through too many phases reduces the comprehension. Similarly, evolutions which are alternations between two phases (e.g., ) hinder the comprehension.
Based on these observations, we assume that the resulting partition must:

regroup observations having similar descriptions into the same cluster (just as traditional clustering does). The clusters represent a certain type of evolution;

create temporally coherent clusters, with limited extent in time. In order for a cluster to be meaningful, it should regroup observations which are temporally close (be contiguous on the temporal dimension). If there are two different periods with similar evolutions (e.g. two economical crises), it is preferable to have them regrouped separately, as they represent two distinct phases. Furthermore, while it is acceptable that some evolutions exist during the entire period, usually the resulted clusters should have a limited temporal extent;

segment, as contiguously as possible, the series of observations for each entity. The sequence of segments will be interpreted as the sequence of phases through which the entity passes.
In this paper, we propose a new timeaware dissimilarity measure that takes into account the temporal dimension. Observations that are close in the description space, but distant in time are considered as dissimilar. We also propose a method to enforce the segmentation contiguity, by introducing a penalty term based on the Normal Distribution Function. We combine the two propositions into a novel timedriven constrained clustering algorithm, TDCKMeans, which creates a partition of coherent clusters, both in the multidimensional space and in the temporal space. This algorithm uses soft semisupervised constraints to encourage adjacent observations belonging to the same entity to be assigned to the same cluster. The proposed algorithm constructs the clusters that serve as evolution phases and segments the observations series for each entity.
The paper is organized as follows. In Section 2 we present some previous related works and, in Section 3, we introduce the temporalaware dissimilarity function, the contiguity penalty, function the TDCKMeans algorithm and the graph structure induction method. In Section 4, we present the dataset that we use, the proposed evaluation measures and the obtained results. Finally, in Section 5, we draw the conclusion and plan some future extensions.
2 Related work
Leveraging partial expert knowledge into clustering represents the domain of semisupervised clustering. The expert knowledge is under the form of either class labels, or pairwise constraints. Pairwise constraints are either “mustlink” (the observations must be placed in the same cluster) or “cannotlink” (the two observations cannot be placed in the same cluster). Depending on the method in which supervision is introduced into clustering, divides the semisupervised clustering methods into two classes: a) the similarityadapting methods , which seek to learn new similarity measures in order to satisfy the constraints, and b) the searchbased methods in which the clustering algorithm itself is modified.
The literature presents some examples of algorithms used to segment a series of observations into continuous chunks. In , the daily tasks of a user are detected by segmenting scenes from the recordings of his activities. Semisupervised mustlink constraints are set between all pairs of observations, and a fixed penalty is inflicted when the following conditions are fulfilled simultaneously: the observations are not assigned to the same cluster and the time difference between their timestamps is less than a certain threshold. A similar technique is used in , where constraints are used to penalize nonsmooth changes (over time) on the assigned clusters. This segmenting technique is used to detect tasks performed during a day, based on video, on sound and on GPS information. In , the objects appearing in an image sequence are detected by using a hierarchical descending clustering, that regroups pixels into large temporally coherent clusters. This method seeks to maximize the cluster size, while guaranteeing intracluster temporal consistency. All of these techniques consider only one series of observations (a single entity) and must be adapted for the case of multiple series. The main problem of a threshold based penalty function is to set the value of the threshold, which is usually datadependent. Optimal matching is used in to discover trajectory models, while studying the destandardization of typical life courses.
The temporal dimension of the data is also used in some other fields of Information Retrieval. In , constrained clustering is used to scope temporal relational facts in the knowledge bases, by exploiting temporal containment, alignment, succession, and mutual exclusion constraints among facts. In , clustering is used to segment temporal observations into continuous chunks, as a preprocessing phase. A graphical model is proposed in , that uses a probabilistic model in which the timestamp is part of the observed variables, and the story is the hidden variable to be inferred. But still, none of these approaches seek to create temporally coherent partitions of the data, mainly using the temporal dimension as a secondary information.
In the following sections, we propose a dissimilarity measure, a penalty function and a clustering algorithm in which the temporal dimension has a central role, and which address the limitations existing in the above presented work.
3 TemporalDriven Constrained Clustering
The observations that need to be structured can be written as triples : . is the vector in the multidimensional description space which describes the entity at the moment of time .
Traditional clustering algorithms input a set of multidimensional vectors, which they regroup in such a way that observations inside a group resemble each other as much as possible, and resemble observations in other groups as little as possible. KMeans is a clustering algorithm based on iterative relocation, that partitions a dataset into clusters, locally minimizing the sum of distances between each data points and its assigned cluster centroids . At each iteration, the objective function
is minimized until it reaches a local optimum.
Such a system is appropriate for constructing partitions based solely on , the description in the multidimensional space. It does not take into account the temporal order of the observations, nor the structure of the dataset, the fact that observations belong to entities. We extend to the temporal case by adding to the centroids a temporal dimension , described in the same temporal space as the observations. Just like its multidimensional description vector , the temporal component does not necessary need to exist in the temporal set of the observation. It is an abstraction of the temporal information in the group, serving as a cluster timestamp. Therefore, a centroid will be the couple .
We propose to adapt the KMeans algorithm to the temporal case by adapting the Euclidean distance, normally used to measure the distance between an element and its centroid. This novel temporalaware dissimilarity measure takes into account both the distance in the multidimensional space and in the temporal space. In order to ensure the temporal contiguity of observations for the entities, we add a penalty whenever two observations that belong to the same entity are assigned to different clusters. The penalty depends on the time difference between the two: the lower the difference, the higher the penalty. We integrate both into the TemporalDriven Constrained KMeans (TDCKMeans), which is a temporal extension of KMeans. TDCKMeans searches to minimize the following objective function:
(1) 
where is our temporalaware (TA) dissimilarity measure (detailed in the next section), is the cost function that determines the penalty of clustering adjacent observations of the same entity into different clusters, and is the set of observations in cluster .
3.1 The temporalaware dissimilarity measure
The proposed temporalaware dissimilarity measure combines the Euclidean distance in the multidimensional space and the distance between the timestamps. We propose to use the following formula:
(2) 
where is the classical norm and and are the diameters of , and respectively (the largest distance encountered between two observations in the multidimensional description space and, respectively, in the temporal space). The following properties are immediate:


and

or
Figure 2 plots the temporalaware dissimilarity measure as a color map, depending on the multidimensional component and the temporal component. The horizontal axis represents the normalized multidimensional distance (). The vertical axis represents the normalized temporal distance (). The blue color shows a temporalaware measure close to the minimum and the red color represents the maximum. The dissimilarity measure is zero if and only if the two observations have equal timestamps and equal multidimensional description vectors. Still, it suffices for only one of the components (temporal, multidimensional) to attend the maximum value for the measure to reach its maximum. The measure behaves similar to a MAX operator, always choosing a value closer to the maximum of the two components. The formula for the temporalaware dissimilarity measure was chosen so that any algorithm that seeks to minimize an objective function based on this measure, will need to minimize both its components. This makes it suitable for algorithms that search to minimize both the multidimensional and the temporal variance in clusters.
Both components that intervene in the measure follow a function like . This function provides a good compromise: it is tolerant for small values of (small time difference, small multidimensional distance), but decreases rapidly when augments. The temporalaware dissimilarity measure is an extension of the Euclidean function. If the timestamps are unknown and set to be all equal, the temporal component is canceled and the temporalaware dissimilarity measure becomes a normalized Euclidean distance. In Section 4.4, we evaluate the behavior of the proposed dissimilarity function. We will call TemporalDriven KMeans the algorithm that is based on the KMeans’ iterative structure and uses the temporalaware dissimilarity measure to asses similarity between observations. Note that TemporalDriven KMeans, relative to TDCKMeans, has no contiguous segmentation penalty function (the contiguous segmentation penalty function is detailed in the next section).
3.2 The contiguity penalty function
The penalty function encourages temporally adjacent observations of the same entity to be assigned to the same cluster. We use the notion of soft pairwise constraints from semisupervised clustering. A “mustlink” soft constraint is added between all pairs of observations belonging to the same entity. The clustering is allowed to break the constraints, while inflicting a penalty for each of these violations. The penalty is more severe if the observations are closer in time. The function is defined as:
(3) 
where is a scaling factor and, at the same time, the maximum value taken by the penalty function; is a parameter which controls the width of the function. is dataset dependent and can be set as a percentage of the average distance between observations. is a function that returns if is true and otherwise.
The function resembles to the positive side of the Normal Distribution function, centered in zero. The function has a particular shape, as represented in Figure 3. For small time differences, it descends very slowly, thus inflicting a high penalty for breaking a constraint. As the time difference increases, the penalty decreases rapidly, converging towards zero. When is small, the functions value descends very quickly with the time difference. The function produces penalties only if the constraint is broken for adjacent observation. For high values of , breaking constraints for distant observations cause high penalties, therefore creating segmentations with large segments. Figure 3 shows the evolution of the penalty function with the time difference between two observations, for multiple values of and for .
An advantage of the proposed function is that it requires no time discretization or setting a fixed window width, as proposed in . The parameter permits the fine tuning of the penalty function. In Section 4.4, we evaluate Constrained KMeans, which is an extension of KMeans, to which we add the proposed contiguity penalty function (but which does not take into account the temporal dimension when measuring the distance between observations). The influence of both and will be studied in Section 4.5.
3.3 The TDCKMeans algorithm
The time dependent distance encourages the decrease of both the temporal and multidimensional variance of clusters; meanwhile the penalty function favors the adjacent observations belonging to the same entity to be assigned to the same cluster. The rest of the TDCKMeans algorithm is similar to the KMeans algorithm. It seeks to minimize by iterating an assignment phase and a centroid update phase until the partition does not change between two iterations. The outline of the algorithm is given in Algorithm 1.
The choose_random function chooses randomly, for each centroid , an observation and sets . In the assignment phase, for every observation , the best_cluster function chooses a cluster so that the temporalaware dissimilarity measure from to the cluster’s centroid , added to the cost of penalties possibly incurred by this cluster assignment, is minimized. It resumes to solving the following equation:
This guaranties that the contribution of to the value of diminishes or stays constant. Overall, this assures that diminishes in the assignment phase (or stays constant).
In the centroid update phase, the update_centroid function recalculates the cluster centroids using the observations in and the assignment at the previous iteration. Therefore the contribution of each cluster to the function is minimized. Each of the temporal and the multidimensional components is calculated individually. In order to find the values that minimize the objective function, we need to solve the equations:
(4) 
By replacing equations (2) and (3) in (1), we obtain the following formula for the objective function:
(5) 
Therefore, from equations (4) and (3.3), we obtain the centroid update formulas:
(6) 
Just like the centroid update phase in KMeans, the new centroids in TDCKMeans are also averages over the observations. Unlike KMeans, the averages are weighted for each component, using the distance from the other. For example, each observation contributes to the multidimensional description of the new centroid, proportional with its temporal centrality in the cluster. Observations that are more distant in time (from the centroid) contribute less to the multidimensional description than the ones being closer in time. A similar logic applies to the temporal component. The consequence is that the new clusters are coherent both in the multidimensional space and in the temporal one.
Algorithm’s complexity
Equation (3.3) shows that TDCKMeans’ complexity is , due to the penalty term. Still, the equation can be rewritten, so that only observations belonging to the same entity are tested. If is the number of entities and is the maximum number of observations associated with each entity, then . The complexity of TDCKMeans is , which is well adapted to Social Science and Humanities datasets, where often a large number of individuals is studied over a relatively short period of time ().
3.4 Finetuning the ratio between components
The temporalaware dissimilarity measure, as presented in Equation (2), gives equal importance to both the multidimensional component and the temporal component. This might pose problems when the data are not uniformly distributed both in the multidimensional descriptive space and in the temporal space. If the medium standard deviation reported to the medium distance between pairs of observations is greater in one space than in the other, giving equal weight to the components can lead to important bias in the clustering process. E.g. observations that are very uniformly distributed in the temporal space (same number of observations for each timestamp) and, at the same time, rather compactly distributed in the description space. In this case, in average, the temporal component weight more in the dissimilarity measure than the multidimensional component. Consequently, the clustering is biased towards the temporal cohesion of clusters. Similarly, in some applications, it is desirable to privilege one component over the other. E.g. on a large enough scale, user roles in social networks have a temporal component (new types of roles might appear over the years). But in a limited time span, it is perfectly acceptable that the roles can coexist simultaneously. Therefore, the temporal component should have only a mild impact on the overall measure.
We adjust the ratio between the two components by using two tuning factors and . weights the multidimensional component of the temporalaware dissimilarity measure, whereas weights the temporal component. Equation (2) can be rewritten as:
(7) 
When the tuning factor for a certain component is set at zero, the respective component does not contribute to the temporalaware measure. When the tuning factor is set to one, no penalty is inflicted to the contribution of the respective component to the measure. It is immediate that equation (2) is a special case of equation (7), with and (no weights).
Setting the weights and
and are not independent one from another, their values are set using a unique parameter .
(8) 
acts as a slider, taking values from to . Figure 4 shows the evolution and with . Also, Figure 5 shows the color map of the temporalaware dissimilarity measure for multiple values of .
When , then and . The multidimensional component is eliminated and only the time difference between the two observations is considered. The temporalaware measure becomes a normalized time difference (). The color map in Figure (a)a () shows that the values of the dissimilarity measure is independent of the multidimensional component.
As the value of increases, the weight of the descriptive component increases as well. In Figure (b)b (), the multidimensional component has a limited impact on the overall measure. When , then and , both components have equal importance, as proposed initially in Equation (2). In Figure (c)c (), the color map shows that the multidimensional component has a larger impact then the temporal component. Large values of the temporal component have only moderate influence over the measure. When (color map in Figure (d)d), then and , the temporal dimension is eliminated and the measure becomes a normalized Euclidean distance ().
Since the temporalaware dissimilarity measure is used in the objective function in Equation (3.3), the later changes accordingly to integrate the tuning factors. and behave as constants in the derivation formulas in Equation (4). As a result, the centroid update formulas in Equation (6) are rewritten as:
The tuning between the multidimensional and temporal component in the temporalaware dissimilarity measure propagates into the centroid update formula of TDCKMeans. We study, in Section 4.6, the influence of the tuning parameter and we propose an heuristic to set its value.
3.5 Inferring a graph structure for the temporal clusters
In Figure (a)a, when discussing the desired output of our system, we presented the obtained temporal clusters under the form of a graph. The nodes represent the evolution phases and an edge between two nodes and indicates that the transition is part of a typical evolution. Since each temporal cluster is interpreted as an evolution phase, the visualization under the form of a graph allows quick understanding of how the different phases are organized both (i) in time (phases on the left side of Figure (a)a have a lower timestamp than those on the right side) and (ii) considering the transitions of the entities through phases. Intuitively, the strength of the connection between two phases is proportional with the number of entities which present transitions between the two given phases.
We consider that an entity presents a transition between and () if and only if two consecutive observations exist, associated with the given entity, where the first observation (ordered by their timestamp) is clustered under and the second one is clustered under . Formally:
(9) 
Furthermore, we define the intersection similarity measure between two phases, which is based on the normalized number of entities that present a transition between the two phases. Formally, we define the :
(10) 
where and needs to be maximized.
We infer a graph structure between the temporal clusters, by constructing an adjacency matrix using the intersection similarity measure. The graph construction is performed a posteriori, after the temporal clusters are calculated. We define the adjacency matrix , where . By replacing equations 9 and 10 into this definition, we obtain:
(11) 
We construct , a binary adjacency matrix, by using a threshold :
The filtering parameter is dataset dependent and automatically setting its value is part of the perspectives of our work.
4 Experiments
4.1 Dataset
Experimentations with TimeDriven Constrained KMeans are performed on a dataset issued from political sciences: Comparative Political Data Set I . It is a collection of political and institutional data, which consists of annual data for 23 democratic countries for the period from 1960 to 2009. The dataset contains 207 political, demographic, social and economic variables.
The dataset was cleaned by removing redundant variables (e.g. country identifier and postal code) and the corpus was preprocessed by removing entity bias from the data. For example, it is difficult to compare, on the raw data, the evolution of population between populous country and one with fewer inhabitants, since any evolution in the 50 years timespan of the dataset will be rendered meaningless by the initial difference. Inspired from panel data econometrics , we remove the entityspecific, timeinvariant effects, since we assume them to be fixed over time. We subtract from each value the average over each attribute and over each entity. We retain the timevariant component, which is in turn normalized, in order to avoid giving too much importance to certain variables. The obtained dataset is under the form of triples .
4.2 Qualitative evaluation
When studying the evolution of countries over the years, it is quite obvious for the human reader why the evolutions of the eastern European countries resemble each other for most of the second half of the twentieth century. The reader would create a group entitled “Communism”, extending from right after the Second World War until roughly 1990, for defining the typical evolution of communist countries. One would expect that, based on a political dataset, the algorithms would succeed in identifying such typical evolutions and segment the time series of each of these countries accordingly. Figure 6 shows the typical evolution patterns constructed by TDCKMeans (with and , obtained as shows in Section 4.5), when asked for 8 clusters. The distribution over time of observations in each cluster is given in Figure (a)a. All constructed clusters are fairly compact in time and have limited temporal extents. They can be divided into two temporal groups. In the first one, clusters to consistently overlap. Same for clusters to , in the second group. This indicates that the evolution of each country passes by at least one cluster from each group. The turning point between the two groups is around 1990. Figure (b)b shows how many countries belong in a certain cluster for each year. Clusters and contain most of the observations, suggesting the general typical evolution.
The meaning of each constructed cluster starts to unravel only when studying the segmentation of countries over clusters, in Figure (c)c. For example, cluster regroups the observations belonging to Spain, Portugal and Greece from 1960 up until around 1975. Historically, this coincides with the nondemocratic regimes in those countries (Franco’s dictatorship in Spain, the “Regime of the Colonels” in Greece). Likewise, cluster contains observations of countries like Denmark, Finland, Iceland, Norway, Sweden and New Zealand. This cluster can be interpreted as the “Swedish Social and Economical Model” of the Nordic countries, to which the algorithm added, interestingly enough, New Zealand. In the second period, cluster regroups observations of Greece, Ireland, Spain, Portugal and Belgium, the countries which seemed the most fragile in the aftermaths of the economical crises of 2008.
Similar conclusions can be drawn from the constructed graph structure, presented in Figure 7. Each temporal cluster is represented as a node and the scores indicated on each edge are calculated as shown in Equation 11. The graph containing all transitions are represented in Figure (a)a (no filtering, ). We obtain a graph containing more general evolutions, by filtering with the threshold . Some “rare” phases completely disappear (i.e., phases and ) together with some of the arcs. We recognize in the resulted graph, shown in Figure (b)b, some of the evolutions identified earlier. The evolution corresponds to the “Swedish Social and Economical Model”, whereas the evolution identifies the fragile European economies of the 2008 economical crises. From the filtered evolution graph, another typical evolution emerges: , which is present for countries as USA, Germany, Italy and France. We interpret this evolution as that of countries with stable social and economical environments.
4.3 Evaluation measures
Since the dataset contains no labels to report to as ground truth, we use the classical Information Theory measures in order to numerically evaluate the proposed algorithms. We evaluate separately each of the three goals that we propose in Section 1.
Create clusters that are coherent in the multidimensional description space. It is desirable that observations that have similar multidimensional descriptions to be partitioned under the same cluster. The similarity in the description space is measured by the multidimensional component of the temporalaware dissimilarity measure. This goal is pursued by all classical clustering algorithms (like KMeans) and any traditional clustering evaluation measure can be used to asses it. We choose the mean cluster variance, which is traditionally used in clustering to quantify the dispersion of observations in clusters. The MDvar measure is defined as:
Create temporally coherent clusters, with limited extend in time. This goal is very similar to the previous one, translated in the temporal space. It is desirable that observations that are assigned to the same cluster to be similar in the temporal space (i.e. to be close in time). The similarity in the temporal space is measured by the temporal component in the temporalaware dissimilarity measure. The limited time extent of a centroid implies small temporal distances between observations timestamp and the centroid timestamp. As a result, the variance can also be used to measure the dispersion of clusters in the temporal space. Similarly to MDvar, the Tvar measure is defined as:
Segment the temporal series of observations of each entity into a relatively small number of contiguous segments. The goal is to have successive observations belonging to an entity grouped together, rather that scattered in different clusters. The Shannon entropy can quantify the number of clusters which regroup the observations of an entity, but it is insensible to alternations between two classes (evolutions like ). We evaluate using an adapted mean Shannon entropy of clusters over entities, which weights the entropy by a penalty factor depending on the number of continuous segments in the series of each entity. The ShaP measure is calculated as:
where is the number of changes in the cluster assignment series of an entity, is the minimal required number of changes and is the number of observations for an entity. For example, in Figure 8, if the series of 11 observations of an entity is assigned to two clusters, but it presents 4 changes, the entropy penalty factor will be . The ShaP score for this segmentation will be , compared to a score of of the “ideal” segmentation (only two contiguous chunks).
The “ideal” values for MDvar, Tvar and ShaP is zero and, in all of the experiments presented in the following sections, we search to minimize the values of the three measures.
4.4 Quantitative evaluation
For each combination of algorithms and parameters, we execute 10 times and compute only the average and the standard deviation. We vary , the number of clusters, from 2 to 36. The performances of five algorithms are compared from a quantitative point of view:

Simple KMeans  clusters the observations based solely on their resemblance in the multidimensional space;

TDCKMeans  the TemporalDriven Constrained Clustering algorithm proposed in Section 3.3. , and ;

tcKMeans  the temporal constrained clustering algorithm proposed in . It uses a threshold penalty function when observations and are not assigned to the same cluster. We adapt it to the multientity case by applying it only to observations belonging to the same entity. Parameters: .
The parameter in tcKMeans should not be mistaken with the parameter in TDCKMeans, as they do not have the same meaning. In tcKMeans, controls the weight of the penalty function, whereas in TDCKMeans is the finetuning parameter.
Algorithm  MDvar  Tvar  ShaP  

Scores  Simple KMeans  120.59  2.97  48.01  8.87  2.15  0.23 
TempDriven KMeans  122.98  2.85  19.97  5.39  2.58  0.18  
Constrained KMeans  132.69  8.07  103.15  42.98  1.24  0.5  
TDCKMeans  127.81  3.96  27.54  5.81  2.06  0.2  
tcKMeans  123,04  3.8  62.44  24.16  1.79  0.32  
% Gain  TempDriven KMeans  1.99%  58.40%  19.63%  
Constrained KMeans  10.04%  114.84%  42.21%  
TDCKMeans  5.99%  42.64%  4.19%  
tcKMeans  2.03%  30.05%  16.99% 
Obtained results.
All the parameters are determined as shown in Section 4.5. Table 1 shows the average values for the indicators, as well as the average standard deviation (in italic) obtained by each algorithm over all values of . The average standard deviation is only used to give an idea of the order of magnitude of the stability of each algorithm. Since Simple KMeans, TemporalDriven KMeans and Constrained KMeans are designed to optimize mainly one component, it is not surprising that they show the best scores for, respectively, the multidimensional variance, the temporal variance and the entropy (best results in boldface). TDCKMeans seeks to provide a compromise, obtaining in two out of three cases the second best score. It is noteworthy that the proposed temporalaware dissimilarity measure used in TemporalDriven KMeans provides the highest stability (the lowest average standard deviation) for all indicators. Meanwhile, the constrained algorithms (Constrained KMeans and tcKMeans) show high instability, especially on Tvar. TDCKMeans shows a very good stability. The second part of Table 1 gives the relative gain of performance of each of the proposed algorithms over Simple KMeans. It is noteworthy the effectiveness of the temporalaware dissimilarity measure proposed in Section 3.1, with a 58% gain of Temporal Variance and less than 2% loss of multidimensional variance. The proposed dissimilarity measure greatly enhances the temporal cohesion of the resulted clusters, without a significant scattering of observations in the multidimensional space. Similarly, the Constrained KM shows an improvement in the contiguity measure ShaP of 42%, while losing 10% multidimensional variance. By comparison, tcKMeans shows modest results, improving ShaP by only 17% and still showing important losses on both Tvar (30%) and MDvar (2%). This proves that the threshold penalty function proposed in literature has lower performances than our newly proposed contiguity penalty function. TDCKMeans combines the advantages of the other two algorithms, providing an important gain of 43% of temporal variance and increasing the ShaP measure by more than 4%. Nonetheless, it shows a 6% loss of MDvar.
Varying the number of clusters
Similar conclusions can be drawn when varying the number of clusters. MDvar (Figure (a)a) decreases, for all algorithms, as the number of cluster increases. It is well known in clustering literature that the intracluster variance decreases steadily with the increase of number of clusters. As the number of clusters augments, so does the differences of TDCKMeans and Constrained KMeans, when compared to the Simple KMeans algorithm. This is due to the fact that the constraints do not let too many clusters to be assigned to the same entity, resulting in the convergence towards a local optimum, with a higher value of MDvar. An opposite behavior is shown by the ShaP measure in Figure (c)c, which increases with the number of clusters. It is interesting to observe how the MDvar and the ShaP measures have almost opposite behaviors. An algorithm that shows the best performances on one of the measures, also shows the worst on the other. The temporal divergence in Figure (b)b shows a very sharp decrease for a low number of clusters, and afterwards remains relatively constant.
4.5 Impact of parameters and
The parameter controls the impact of the contiguity constraints in Equation (3). When set to zero, no constraints are imposed, and the algorithm behaves just like the Simple KMeans. The higher the values of , the higher the penalty inflicted when breaking a constraint. When is set to large values, the penalty factor will take precedence over the similarity measure in the objective function. Observations that belong to a certain entity will be assigned to the same cluster, regardless of their resemblance in the description space. When this happens, the algorithm cannot create partitions with higher number of clusters than the number of entities. In order to evaluate the influence of parameter , we execute the Constrained KMeans algorithm with varying from 0 to 0.017 with a step of 0.0005. The value of is set at 3, and 5 clusters are constructed. For each value of , we executed 10 times the algorithm and we plot the average obtained values. Figure (a)a shows the evolution of measures MDvar and ShaP with . When both MDvar and ShaP have the same values as for Simple KMeans. As increases, so does the penalty for noncontiguous segmentation of entities. MDvar starts to increase rapidly, while ShaP decreases rapidly. Once reaches higher values, the measures continue their evolution, but with a leaner slope. In the extreme case, in which all observations are assigned to the same cluster regardless of their similarity, the ShaP measure will reach zero.
The parameter controls the width of the penalty function in Equation (3). As Figure 3 shows, when has a low value, a penalty is inflicted only if the time difference of a pair of observations is small. As the time difference increases, the function quickly converges to zero. As increases, the function decreases with a leaner slope, thus also taking into account observations which are farther away in time. In order to analyze the behavior of the penalty function when varying , we have executed the Constrained KMeans, with ranging from 0.1 to 8, using a step of 0.1. was set at 0.003 and 10 clusters were constructed. Figure (b)b plots the evolution of measures MDvar and ShaP with . The contiguity measure ShaP decreases almost linearly as increases, as the series of observations belonging to each entity gets segmented in larger chunks. At the same time, the multidimensional variance MDvar increases linearly with . Clusters become more heterogeneous and variance increases, as observations get assigned to clusters based on their membership to an entity, rather than their descriptive similarities.
Varying and for the tcKMeans proposed in yields similar results, with the MDvar augmenting and the ShaP descending, when and increase. For the tcKMeans, these evolutions are linear, whereas for the Constrained KMeans they are exponential, following a trend line of function . Plotting the evolution of the MDvar and the ShaP indicators on the same graphic, provides a heuristic for choosing the optimum values for the parameters of the Constrained KMeans and the TDCKMeans, respectively the parameters of the tcKMeans. Both curves are plotted with the vertical axis scaled to the interval . Their point of intersection determines the values of the parameters (as shown in Figure (a)a and (b)b). The disadvantage of such a heuristic would be that a large number of executions must be performed with multiple values for the parameters before the “optimum” can be found.
4.6 The tuning parameter
The parameter , proposed in Section 3.4, allows the fine tuning of the ratio between the multidimensional component and the temporal component in the temporalaware dissimilarity measure. When is close to 1, the temporal component is predominant. Conversely, when is close to 1, the multidimensional component takes precedence. The two components have equal weights when . To evaluate the influence of parameter , we executed TemporalDriven KMeans with varying from 1 to 1 with a step of 0.1. In order not to bias the results and to evaluate only the impact of the tuning parameter, we remove the contiguity constraints from the objective function , by setting . For each value of , we executed 10 times and we present the average values.
Figure (a)a shows the evolution of measures MDvar and Tvar with . For low values of , the value of the temporalaware dissimilarity measure is given mainly by the temporal component, so Tvar shows its lowest value, while MDvar presents its maximum. As increases, MDvar decreases as more importance is given to the multidimensional component. For , the importance of the temporal component remains intact, the increase of Tvar is solely the result of the algorithm converging to a local optimum which also takes into account the multidimensional component. For , the impact of the multidimensional component stays constant, whereas the importance of the temporal components diminishes. As a result, MDvar continues its decrease and Tvar increases sharply. For the temporal component is basically ignored from the measure. The TemporalDriven KMeans behaves just like Simple KMeans. Figure (b)b shows the evolution of ShaP alongside MDvar. Even if the contiguity penalty component was neutralized by setting , the value of ShaP is not constant, but it descends with . For low values of , the temporal component is predominant in the similarity measure. This generates partitions where every cluster regroups all observations from a specific period, regardless of their multidimensional description. This means that all entities have segments in all the clusters, which leads to a high value of ShaP.
It is noteworthy that the evolution of the indicators is not linear with . As increases, Tvar augments only very slowly and picks up the pace only for large values of . This indicates that the temporal component has an inherent advantage over the multidimensional one. As we presumed in Section 3.4, this is due to the intrinsic nature of the dataset and the main reason why the tuning parameter was introduced. The distributions of observations in the multidimensional and temporal spaces is different: in the temporal space, the observations tend to be evenly distributed, whereas in the multidimensional description space, they cluster together. To quantify this, we calculate the ratio between average standard deviation and average distances between pairs of observations:
where is replaced with or ( the descriptive or the temporal dimension). On Comparative Political Data Set I, and . This shows that observations are a lot more dispersed in the temporal space than in the multidimensional description space. This explains why Tvar augments very slowly with and starts to increase more rapidly only starting from .
Following the heuristic proposed in Section 4.5, we can determine a tradeoff value for . As shown in Figure 11, all vertical axes are magnified between their functions’ minimum and maximum values. The tradeoff value for is found at the intersection point of MDvar and Tvar (and MDvar and Shap). This value is set around , showing the dataset’s bias towards the temporal component. This technique for setting the value of the tunning parameter is just a heuristic, the actual value of is dependent on the dataset. This is why we are currently working on a method, inspired from multiobjective optimization using evolutionary algorithms to automatically determine the values of , as well as the other parameters of TDCKMeans (, and ).
5 Conclusion and future work
In this paper, we have studied the detection of typical evolutions from a collection of observations. We have proposed a novel way to introduce temporal information directly into the dissimilarity measure, weighting the Euclidean component in the description space by the temporal component. We have proposed TDCKMeans, an extension of KMeans, which uses the temporalaware dissimilarity measure and a new objective function which takes into consideration the temporal dimension. We use a penalty factor to make sure that the observation series related to an entity get segmented into continuous chunks. We infer a new centroid update formula, where elements distant in time contribute less to the centroid than the temporally close ones. We have proposed an intersection similarity measure between two temporal clusters and a method to calculate a posteriori an adjacency matrix. We use this adjacency matrix in order to structure the detected evolution phases as a graph. From a qualitative point of view, we have shown that our algorithm is capable of detecting comprehensible evolutions based on a Political Science dataset. Quantitatively, we have shown that our proposition consistently improves temporal variance, without any significant losses in the multidimensional variance.
We are currently experimenting with applying the algorithm to other applications, e.g., detection of social roles in social networks, by passing through temporal behavioral roles. A social role is defined as a typical succession of behavioral roles. In our current work, we have inferred a temporal cluster graph structure a posteriori to the construction of the clusters. We have ongoing work toward incorporating the graph construction into the TDCKMeans algorithm, by modifying the objective function in order to take into account the intersection similarity measure and a temporal distance. Another direction of research will be describing the clusters in a human readable form. We work on means to provide them with an easily comprehensible description by introducing temporal information into an unsupervised feature construction algorithm. We are also experimenting a method for setting automatically the values of TDCKMeans’s parameters (, , and ), by using an approach inspired from multiobjective optimization using evolutionary algorithms .
References
 1. Klaus Armingeon, David Weisstanner, Sarah Engler, Panajotis Potolidis, Marlène Gerber, and Philipp Leimgruber. Comparative political data set 19602009. Institute of Political Science, University of Berne., 2011.
 2. Sugato Basu, Arindam Banerjee, and Raymond J. Mooney. Semisupervised clustering by seeding. In International Conference on Machine Learning, pages 19–26, 2002.
 3. Mikhail Bilenko and Raymond J. Mooney. Adaptive duplicate detection using learnable string similarity measures. In Knowledge Discovery and Data Mining, Proceedings of the ninth ACM SIGKDD international conference on, pages 39–48. ACM New York, NY, USA, 2003.
 4. Shixi Chen, Haixun Wang, and Shuigeng Zhou. Concept clustering of evolving data. In Data Engineering, 2009. ICDE’09. IEEE 25th International Conference on, pages 1327–1330, 2009.
 5. David Cohn, Rich Caruana, and Andrew McCallum. Semisupervised clustering with user feedback. In Constrained Clustering: Advances in Algorithms, Theory, and Applications, volume 4, pages 17–32. Cornell University, 2003.
 6. Fernando De la Torre and Carlos Agell. Multimodal diaries. In Multimedia and Expo, 2007 IEEE International Conference on, pages 839–842. IEEE, 2007.
 7. Ayhan Demiriz, Kristin Bennett, and Mark J. Embrechts. Semisupervised clustering using genetic algorithms. In Artificial Neural Networks in Engineering, pages 809–814, 1999.
 8. Brigitte Dormont. Petite apologie des données de panel. Économie & prévision, 87(1):19–32, 1989.
 9. Nizar Grira, Michel Crucianu, and Nozha Boujemaa. Unsupervised and semisupervised clustering: a brief survey. Technical report, A Review of Machine Learning Techniques for Processing Multimedia Content, Report of the MUSCLE European Network of Excellence (FP6), 2005.
 10. Maria Halkidi, Yannis Batistakis, and Michalis Vazirgiannis. On clustering validation techniques. Journal of Intelligent Information Systems, 17(2):107–145, 2001.
 11. Dan Klein, Sepandar D. Kamvar, and Christopher D. Manning. From instancelevel constraints to spacelevel constraints: Making the most of prior knowledge in data clustering. In International Conference on Machine Learning, pages 307–314, 2002.
 12. WeiHao Lin and Er Hauptmann. Structuring continuous video recordings of everyday life using timeconstrained clustering. In IS&T/SPIE Symposium on Electronic Imaging, 2006.
 13. James MacQueen. Some methods for classification and analysis of multivariate observations. In L. M. Cam and J. Neyman, editors, Berkeley Symposium on Mathematical Statistics and Probability, Proceedings of the Fifth, volume 1, pages 281–297, 1967.
 14. Arun Qamra, Belle Tseng, and Edward Y. Chang. Mining blog stories using communitybased and temporal clustering. In Information and Knowledge Management, Proceedings of the 15th ACM international conference on, pages 58–67, New York, NY, USA, 2006. ACM.
 15. Brandon C. S. Sanders and Rahul Sukthankar. Unsupervised discovery of objects using temporal coherence. Technical report, CVPR Technical Sketch, 2001.
 16. Partha Pratim Talukdar, Derry Wijaya, and Tom Mitchell. Coupled temporal scoping of relational facts. In WSDM, pages 73–82, 2012.
 17. Kiri Wagstaff, Claire Cardie, Seth Rogers, and Stefan Schroedl. Constrained kmeans clustering with background knowledge. In International Conference on Machine Learning, Proceedings of the Eighteenth, pages 577–584. Morgan Kaufmann, 2001.
 18. Eric D. Widmer and Gilbert Ritschard. The destandardization of the life course: Are men and women equal? Advances in Life Course Research, 14(12):28–39, 2009.
 19. Eric P. Xing, Andrew Y. Ng, Michael I. Jordan, and Stuart Russell. Distance metric learning with application to clustering with sideinformation. Advances in Neural Information Processing Systems, 15:505–512, 2002.
 20. Qingfu Zhang and Hui Li. Moea/d: A multiobjective evolutionary algorithm based on decomposition. Evolutionary Computation, IEEE Transactions on, 11(6):712–731, 2007.