1 Introduction
In the age of big data, applications deal with data collected at very finegrained time intervals. In many real world applications, the data collected spans long periods of time and can be extremely sparse. For instance, a timeevolving social network that records interactions of users every second results in a very sparse adjacency matrix if observed at that granularity. Similarly, in spatiotemporal data, if one considers GPS data over time, discretizing GPS coordinates based on the observed granularity can lead to very sparse data which may not contain any visible and useful structure. How can we find meaningful and actionable structure in these types of data? A great deal of such datasets are multiaspect in nature and hence can be modeled using tensors. For instance, a threemode tensor can represent a timeevolving graph capturing useruser interactions over a period of time, measuring crime incidents in a city community area over a period of time [1, 17], or measuring traffic patterns [21]. Tensor decomposition has been used in order to extract hidden patterns from such multiaspect data [16, 13, 10]. However the degree of sparsity in the tensor, which is a function of the granularity in which the tensor is formed, significantly affects the ability of the decomposition to discover “meaningful” structure in the data.
Consider a dataset which can be modeled as threemode tensor, where the third mode is temporal as shown in Figure 1. If the granularity of the temporal mode is too fine (in milliseconds or seconds), one might end up with a tensor that is extremely long on the time mode and where each instance of time has very small number of entries. This results in a extremely sparse tensor, which typically is of very high rank, and which usually has no underlying exploitable structure for widely popular and successful tensor decomposition algorithms [16, 13, 10]. However, as we aggregate data points over time, exploitable structure starts to appear (whereby “exploitable” we define the kind of lowrank structure that a tensor decomposition can successfully model and extract). In this paper we set out to explore what is the best such datadriven aggregation of a tensor which leads to better, exploitable, and interpretable structure, and how this fares against the traditional alternative of selecting a fixed interval for aggregation.
As far as tackling the problem above, there is considerable amount of work that focuses on a special case, that of aggregating edges of a time evolving graph into “mature” adjacency matrices based on certain graph properties [18, 19, 20]. In our work, however, we address the problem in more general terms, where the underlying data can be any point process that is observed over time and/or space, and where the aggregation/discretization of the corresponding dimensions directly affects our ability to extract interpretable patterns via tensor decomposition. Effectively, as shown in Figure 1, in this paper we work towards automating the data aggregation starting from raw data into a wellstructured tensor.
Our contributions in this work are as follows:

Novel Problem Formulation: We formally define the problem of optimally aggregating a tensor, which is formed from raw sparse data in their original level of aggregation, into a tensor with exploitable and interpretable structure. We further show that solving this problem optimally is computationally intractable. To the best of our knowledge, this paper is the first to tackle this problem in its general form, and we view our formulation as the first step towards automating the process of creating wellbehaved tensor datasets.

Practical Algorithm: We propose a practical, efficient, and effective algorithm that is able to produce highquality tensors from raw data without incurring the combinatorial cost of the optimal solution. Our proposed method follows a greedy approach, where at each step we decide whether different “slices” of the tensor are aggregated based on a variety of intuitive functions that characterize the “goodness of structure” locally.

Experimental Evaluation: We extensively evaluate our proposed method on synthetic and semisynthetic data, where ground truth is known, and in real data where we use popular heuristic measures of structure goodness to measure success. Furthermore, we conduct a data mining case study on a large real dataset of crime over time in Chicago, where we identify interpretable hidden patterns in multiple time resolutions.

Reproducibility: We make our implementation publicly available at https://github.com/ravdeep003/adaptivegranularitytensors in order to encourage reproducibility of our results.
2 Problem Formulation
2.1 Tensor Definition and Notations
Tensors are multidimensional extensions of matrices, and tensor decompositions are a class of methods that extract latent structure from tensor datasets by extending techniques such as Principal Component Analysis and Singular Value Decomposition. The different “dimensions” of a tensor are usually referred to as “modes”. In this paper, we focus on the CANDECOMP/PARAFAC (henceforth refered to as CP for brevity) decomposition
[9], which is the “rank decomposition” of a tensor, i.e., the decomposition of an arbitrary tensor into a sum of rankone tensors. Mathematically, for a threemode tensor , the CP decomposition is , where is the generalized outer product. Matrices are called “factor matrices”, and each column corresponds to a latent pattern, directly relating an entity of the corresponding mode to a value that can be roughly construed as a soft clustering coefficient [5]. CP has arguably been the most popular tensor decomposition model in applications where the interest is to extract interpretable patterns for exploratory analysis, and thus, we adopt this decomposition model as our standard in this work. In the interest of space, we refer the reader to a number of available surveys [16, 13, 10]. We denote tensors as and matrices as , and we adopt Matlablike notation for indexing.2.2 Tensor decomposition quality
Unsupervised tensor decomposition, albeit very popular, poses a significant challenge: how can we tell whether a computed decomposition is of “high quality”, and how can we go about defining “quality” in a meaningful way? Unfortunately, this happens to be a very hard problem to solve [14], and defining a new measure of quality is beyond the scope of this paper. However, there has been significant amount of work in that direction, which basically boils down to 1) modelbased measures, where the quality is measured by how well a given decomposition represents the intrinsic hidden structure of the data, and 2) extrinsic measures, where the quality is measured by how well the computed decomposition factors perform in a predictive task.
In terms of modelbased measures, the most straightforward one is the fit, i.e., how well does the decomposition approximate the data under the chosen loss function, in a
low rank. Low rank is key, because the number of components (rank) has to be as small and compact as possible in order to lend itself to human evaluation and exploratory analysis. However, fit has been shown to be unstable and prone to errors especially in real and noisy data, thus the community has collectively turned its attention to more robust measures such as the Core Consistency Diagnostic (CORCONDIA for short) [3], which measures how well the computed factors obey the CP model.On the other end of the spectrum, extrinsic quality measures are always tied to a predictive task. A popular such task is community detection, where the tensor consists of a multiview or timeevolving graph, and the frontal slices contain adjacency matrices of that given graph; the task is to use the computed factors for the nodes as features for assigning them to a community label and subsequently measure the quality of that assignment with measures such as the Normalized Mutual Information (NMI) [4, 8, 7].
Both types of quality measure are heuristic and capture different elements of what an enduser would deem good in a set of decomposition factors. In this paper, we are going to use such popular measures of quality in order to characterize the quality of a given tensor dataset . In order to do so, we assume that we have a function which, optimizes the heuristic quality measure for a given tensor over all possible decomposition ranks ^{1}^{1}1In practice, this is done over a small number of low ranks, since lowrank structure is desirable., i.e.,
where are the column factor matrices for . Finally, a useful operation is the mode product, where a matrix is multiplied by the th mode of a tensor (predicated on matching dimensions in the th mode of the tensor and the rows of the matrix), denoted as . For instance, an tensor where and of size , the product multiplies all third mode slices of with and results in a tensor.
2.3 The Trapped Under Ice problem
To give reader an intuition of the problem, consider an example of timeevolving graph which captures social activity over the span of some time. This example can be modeled as threemode tensor of dimensions where “sender” and “receiver” are the first two modes, “time” being the third mode, and nonzero entry in the tensor represents communications between user at a particular time. If the time granularity is extremely finegrained (milliseconds or seconds), there might be only handful of data points at a particular time stamp causing resulting tensor to be extremely sparse and to have a high tensorrank as a result. In that case, might not have any interpretable lowrank structure that can be exploited by CP. In this example we assume that the third mode (time mode) is too finegrained but in reality any mode (one or more) can be extremely fine grained. For example, in spatiotemporal data, where the first two modes are latitude and longitude and the third mode is time, all three modes can suffer from the same problem.
Given tensor which is created using the “raw” granularities, how does one find a tensor (say ) which has better exploitable structure and hence can be decomposed into meaningful latent structure. This, is informally the Trapped Under Ice problem that we define here (which draws an analogy between the good structure that may exist within the data as being trapped under the ice and not visible by mere inspection). Trapped Under Ice has an inherent assumption that the mode in which we aggregate is ordered (e.g., representing time or space), thus permuting the third mode will lead to a different instance of the problem.
More formally we define our problem as follows:
[linecolor=red!60!black,backgroundcolor=gray!20,linewidth=2pt,topline=false,rightline=false, leftline=false]
Given a tensor of dimensions
Find:
A tensor of dimensions with such that
where is a measure of goodness and if slice in tensor is aggregated into slice in the resulting tensor, otherwise .
At first glance, Trapped Under Ice might look like a problem amenable to dynamic programming, since it exhibits the optimal substructure property. However, it lacks the overlapping subproblems property: there are overlapping subproblems across the set of different matrices (e.g., two different matrices may have overlapping subproblems) but not within any single . Thus, we still have to iterate over ’s refer subsection 2.4 for more details.
Structure of : The matrix has a special structure. Here we provide an example. Consider a threemode tensor of dimensions , with the third mode being the time mode. Suppose that the optimal level of aggregation for is .
In this case, is of size and an example of such matrix is:
This aggregates first three slice of to form first slice of , next three to form the second slice and last four to form the third slice. No two matrices will produce the same aggregation. They can have the same but order of aggregation of slices will be different.
2.4 Solving Trapped Under Ice optimally is hard
Solving Trapped Under Ice optimally poses a number of hurdles. First and foremost, the hardness of the problem depends on the definition of function , and most reasonable and intuitive such definitions are very hard to optimize since they are nondifferentiable, noncontinuous, and not concave. So far, in the literature, to the best of our knowledge, there are only heuristics for this quality function. Even so, those heuristic functions can only be evaluated on a single already fullyaggregated tensor, not a partially aggregated version thereof. Thus, Trapped Under Ice can be only solved optimally via enumerating all admissible solutions and choosing the best. In order to conduct this enumeration, we need to calculate the cardinality of the set of all for a given instance of the problem. For an instance of a problem with initial slices, the cardinality of the set of all is To get aggregated slices there are ways to choose each of them leading to a different . This is a number of ways that partition slots can be filled partitioned by blocks. In order to get the final number, we need to sum up over all potential :
Direct corollary of the above lemma is that solving optimally Trapped Under Ice requires calling the function times, which is computationally intractable. There may be small room for improvement by exploiting special structure in the set of all , however, given discontinuities in our objective function , this is not be a feasible alternative either. In this paper we define proxy quality functions that lend themselves to partial evaluation on a partially aggregated solution, thus allowing for efficient algorithms Thus, in the next section we propose a fast greedy approach which locally optimizes different criteria quality.
3 Proposed Method
In this section, we propose our efficient and effective greedy algorithm called IceBreaker which takes a tensor as an input, which has been created directly from raw data, and has no exploitable structure. and returns a tensor which maximizes the interpretable and exploitable structure. The basic idea behind IceBreaker is to make a linear pass on the mode for which the granularity is suboptimal, and using a number of intuitive and locally optimal criteria for goodness of structure (henceforth referred to as utility functions), we greedily decide whether a particular slice across that mode needs to be aggregated^{2}^{2}2For the purposes of our work, we use matrix addition as aggregation of slice but this might not be the case and would depend on the problem domain. Other aggregation functions that can be used are OR, min, max, depending on the application domain (e.g., binary data). into an existing slice or contains goodenough structure to stand on its own. IceBreaker can choose from a number of intuitive utility functions which are based on different definitions of good quality in matrices and graphs (in cases where we are dealing with underlying graph data).
3.1 The IceBreaker algorithm
Algorithm 1 gives a high level overview of IceBreaker . More specifically, the algorithm takes a threemode tensor of dimension as an input and loops over all the slices of tensor . Two slices next to each other get aggregated into a single slice if a certain utility function has stabilized, i.e., if aggregating the two slices does not offer any additional utility (larger than a particular threshold), then the second slice should not be aggregated with the first, and should mark the beginning of a new slice.
Consider a threemode tensor with time as third mode of dimension is ran through IceBreaker with a particular utility function. Our algorithm iterates over the time mode ( slices) and aggregates slices as decided by the utility function. IceBreaker is agnostic to utility function used. Let us consider a slice that has been aggregated into a single slice from indices to called previous slice and another aggregated slice from indices to called a candidate slice. Both previous and candidate slice are passed to utility function separately to obtain a value each called previous and current value respectively. These values are compared (line in algorithm 1) to decide whether slice is absorbed(line in algorithm 1) into previous slice or previous slice has stabilized and entry is added in to indicate which indices of tensor are aggregated together(line in algorithm 1). Now slice becomes the previous slice and aggregated slice of and become the candidate slice, the whole process is repeated until all the slices are exhausted.
Note that IceBreaker ’s complexity is linear in terms of the slices of the original tensor, and its overall complexity depends on the specific utility function used (which is called times).
3.1.1 Utility functions:
In the subsection, we summarize a number of intuitive utility functions that we are using in this work. This list is by no means exhaustive, and can be augmented by different functions (or function combinations) that capture different elements of what is good structure and can be informed by domainspecific insights.

Norm: We use multiple norm types to find adaptive granularity of a tensor. For a given threshold, if rate of change of norm between previous and candidate slice is less than the threshold, candidate slice is not selected. Our assumption in this case is no significant amount of information is being added to previous slice and is considered to have been stabilized. Matrix is updated accordingly with indices of the previous slice (aggregated slices in previous slice). Otherwise the candidate slice is selected and the process continues until all the slices are exhausted. Different norms demonstrated in this work are Frobenius, 2norm, and Infinity norm.

Matrix Rank: In case of matrix rank, we focus on the reconstruction rank, which is typically much lower than the full rank of the data, but captures the essence of the number of components within the slice. In this case, we consider previous slice to be stabilized if the matrixrank of previous slice decreases by addition of new slice, no more slices are added and an entry in matrix is added. We keep aggregating slices if the matrixrank of the slice is increasing or remains constant.

Missing Value Prediction: If a piece of data has good structure, when we hide a small random subset of the data, the remaining data can successfully reconstruct the hidden values, under a particular model that we have chosen. To this end, we employ a variant of matrix factorization based collaborative filtering [11]
as a utility function to see how good is the aggregated matrix in predicting certain percent of missing values. This utility function takes percent of missing value as a parameter, hides those percent of non zeros values in the matrix. Our implementation of matrix factorization with Stochastic Gradient Descent tries to minimize the loss function:
where is a given slice, are factor matrices for a given rank (typically chosen using the same criterion as the matrix rank above), and is the set of observed (i.e., nonmissing) values. In order to create a balanced problem, since we are dealing with very sparse slices, we conduct negative sampling where we randomly sample as many zero entries as there are nonzeros in the slice, and this ends up being the set of observed values. 
Graph Properties: There has been a significant amount of work in graph mining with respect to aggregation of temporal graph [18, 19, 20], taking inspiration from [18], we use the following functions:

Average degree: Similar to norm, we consider a previous slice to stabilized if the rate of change of average degree between previous and candidate slice is less than the threshold.

Connected components: We consider number of connected components that has more that one node. If the number of connected components remain the same we keep aggregating the slices and if the increases we consider it to stabilize.

4 Experimental Evaluation
In this section we present a thorough evaluation of IceBreaker using variety of data, including synthetic and semisynthetic data (where ground truth is known), and real data where there is no ground truth, and where we empirically evaluate our analysis using a number of criteria described in detail below. We implement our method in Matlab using tensor toolbox library [2]. For small variations in parameters to the utility functions we did not observe much deviations however we plan to investigate in greater detail the effects of various thresholds in an extended version.
4.1 Evaluation measures
When formulating the problem, we did not specify a quality function to be maximized, nor did we use such a function in our proposed method. The reason for that is because we reserve the use of different quality functions as a form of evaluation. In particular, we use the two following notions of quality:

Community detection NMI
: We use NMI as a measure of predictive accuracy when the ground truth for the timeevolving data is available. To compute NMI score we perform CP decomposition on the resulting tensor with tensor rank provided by Autoten from above measure and then we use Kmeans clustering on the relevant factor matrices with number of communities in ground truth
[6]. With that each node (row in the factor matrix) gets assigned to a single community we pass this result with the actual ground truth to NMI function to get NMI score.
We should note at this point that the two quality measures above are far from continuous and monotonic functions, thus we do not expect that our IceBreaker progresses the quality will monotonically increase. Thus, we calculate the quality for the final solution of IceBreaker , and we reserve investigating whether monotonic and wellbehaved quality functions exist for future work.
4.2 Baseline methods
A naive way to find tensor can be by aggregating time mode based on some fixed intervals. If time granularity was in milliseconds, then combining one thousand slices to form slices of seconds granularity reducing the third dimension of tensor from to . This can be applied incrementally from seconds to minutes and so on to find a tensor which has some exploitable structure. We compare the resulting tensor determined by IceBreaker against tensors constructed with fixed aggregations. For fixed aggregation we aggregate the temporal with window size of and .
4.3 Performance for synthetic data with ground truth
Creating synthetic data: In order to create synthetic data we follow a twostep process: 1) we use an existing timeevolving tensor generator, proposed in [15], which creates a tensor that has a certain number of latent factors which appear, disappear, and reappear randomly over time, 2) subsequently, we simulate the Trapped Under Ice problem by taking every nonzero entry of that tensor and creating a new slice containing that entry (effectively setting equal to the number of nonzero entries of the tensor). Furthermore, we modify the factor matrices of the first two modes, so that each row has only one nonzero entry, corresponding to a particular “cluster” it belongs to, thus enabling us to have ground truth labels for the first two modes of the data. We create two synthetic scenarios: 1) one where changes in the number of components happen on a fixed time scale (which is ideal for fixed aggregation) and 2) one where changes happen at randomized time windows, which is a more challenging case.
Results for synthetic data: In order to evaluate the performance of IceBreaker , we measure CORCONDIA and NMI on the two synthetic datasets, over 10 different runs. The leftmost^{3}^{3}3Dotted line in the leftmost part of figures 2, 3, 4, 5 and Fig. 6 are average value of the score of that experiment. part of Figure 2 shows the results for the first dataset, where fixed aggregations outperform by a small margin the proposed method, a behavior which was expected, since the dataset has been created with a natural fixed size window of aggregation. The leftmost part of Figure 3 shows the results in the more challenging case, where the natural window of aggregation is randomized and variable. In this case, our proposed method works on par with the fixed aggregation methods. In both scenarios fixed aggregations performed well because of the nature of synthetic data which has very well defined latent structures repeated over period of time.
4.4 Performance for semisynthetic data with ground truth
Creating semisynthetic data: In addition to fully synthetic data, we also create semisynthetic data, where the first two modes come from a realworld graph for which the ground truth labels are given, and the time mode is simulated according to the synthetic data generator used in [7]. The process followed is the same as in the synthetic case, where we take the nonzero elements and create a tensor with equal number of slices. The realworld graph we use for semisynthetic data generation is the American Football used in [7]. Other dataset used is European Email from [12] where we create a tensor such that each nonzero entry corresponds to a slice.
4.5 Different local optima hint to multiple resolutions in time
IceBreaker , depending on what utility function it uses, converges to a different (locally) optimal solution, and all such solutions achieve roughly the same combined quality measure (ratio of CORCONDIA and NMI in this case). Most interestingly, those different local optima actually pertain to different levels of resolution for the aggregation. In the rightmost parts of Figures 2, 3, 4, 5 and Figure 6, we plot the degree of aggregation (measured by the ratio of ) versus the quality measure. We observe distinct clusters of solutions that achieve roughly the same quality, for vastly different aggregation levels, which points to the fact that those local optima reveal a multiresolution view of the raw data.
4.6 Data mining case study
Chicago crime dataset: For our case study we use a dataset provided by the city of Chicago that records different types of crime committed in different areas of the city over a period of time [1, 17]. The tensor we create has modes (area, crime, timestamp), where “community area” and “crime” are discretized by the city of Chicago and “timestamp” is the coarsely aggregated (hourly) timestamp. The dates that we focused on span a period of years, between December 13, 2010 to December 11, 2017.
We ran IceBreaker on the dataset, and in Figure 6 we show its CORCONDIA for different degrees of aggregation () and we observe that all utility functions offer highquality solutions, and those solutions offer a number of different resolutions for analyzing the data. Given those two different resolutions, we decided to drill down and look into the actual tensor components that can be extracted from those different tensors. In the interest of space, we took the tensor created by “Frobeniusnorm” which yields a higher resolution tensor , and the tensor created by “Rank change” which is of low resolution. Tensor contains three highquality components, whereas contains four. Figure 7 shows the two different sets of patterns^{4}^{4}4We omit plotting the temporal mode since we lack external information that we can potentially correlate it with, however, an analyst with such side information can find the different time resolutions of and useful.: interestingly, factor 2 of and factor 3 of pertain to the same spatial and criminal pattern, and the same holds for factors 3 and 4 in and respectively. In summary, tensors and capture similar interpretable patterns over different temporal resolutions, which IceBreaker can successfully discover.
Comparison against fixed aggregation: A natural question is whether the results are qualitatively “better” than the ones by a fixed aggregation. Answering this question heavily depends on the application at hand, however, here we attempt to quantify this in the following way: intuitively, a good set of components offers more diversity
in how much of the data it covers. For instance, a practitioner would prefer a set of results for the Chicago crime dataset where the components span most of the regions of the city and uncover diverse patterns of crime, over a set of components that seem to uncover a particular type of crime. Even though there may be a number of confounding factors, aggregating on a regular time interval may be very good in capturing periodic activity (in this example, crime that exhibits normal periodicity that happens to coincide with the aggregation resolution we have chosen), whereas aggregating adaptively may help discover structure that is more erratic and more surprising. In order to capture this and test this hypothesis, we compute the coverage of entities for the first and second mode of the tensor (i.e., areas of Chicago and crime types in this example) in all the discovered components: for each component, we measure the topk entities, and through that we compute the empirical probability distribution of all entities in the results. A more preferable set of results will have a higher coverage, resulting in a distribution with higher entropy. In Table
1 we show the entropy for both modes 1 and 2 for IceBreaker and for the different fixed aggregations (averaged over 10 different runs), where IceBreaker overall offers more diverse patterns in both space and criminal activity.Frobenius  2Norm  Infinity  Rank  Missing Value  Average  Fixed  Fixed  Fixed  
Norm  Norm  Change  Prediction  Degree  Interval10  Interval100  Interval1000  
Area  2.7699  2.7255  2.7255  2.8554  2.9255  2.7255  2.7255  2.8283  2.5850 
Crime  2.9477  2.9477  2.6416  2.9554  2.8726  2.7255  2.6416  2.7617  2.2516 
4.7 Scalability
IceBreaker is computationally efficient and practical, since it runs it scales linearly as a function of the number of slices in the raw data, as evidenced by Figure 8, which shows runtimes on the Chicago dataset. The actual runtime depends on the cost of each utility function.
5 Related Work
To the best of our knowledge, this is the first attempt at formalizing and solving this problem, especially as it pertains in the tensor and multiaspect data mining domain. Nevertheless, there has been significant amount of work on temporal aggregations in graphs [18, 19, 20] and in finding communities in temporal graphs [7]. The closest work to ours is [18], in which the authors looks at aggregating stream of temporal edges to produce sequence of structurally mature graphs based on a variety of network properties.
6 Conclusions
In this paper we are, to the best of our knowledge, the first to define and formalize the Trapped Under Ice problem in constructing a tensor from raw sparse data. We demonstrate that an optimal solution is intractable and subsequently proposed IceBreaker , a practical solution that is able to identify good tensor structure from raw data, and construct tensors from the same dataset that pertain to multiple resolutions. Our experiments demonstrate the merit of IceBreaker in discovering useful and highquality structure, as well as providing tools to data analysts in automatically extracting multiresolution patterns from raw multiaspect data. In future work we will work towards extending IceBreaker in cases where more than one modes is Trapped Under Ice (naively one can apply IceBreaker
to each mode sequentially, but this disregards joint variation across modes), and extend
IceBreaker for higherorder tensors.References
 [1] Chicago data portal.
 [2] B. Bader and T. Kolda. Matlab tensor toolbox version 2.2. Albuquerque, NM, USA: Sandia National Laboratories, 2007.
 [3] R. Bro and H. A. Kiers. A new efficient method for determining the number of components in parafac models. Journal of chemometrics, 17(5):274–286, 2003.
 [4] Evangelos E. Papalexakis, L. Akoglu, and D. Ienco. Do more views of a graph help? community detection and clustering in multigraphs. In IEEE FUSION’13.
 [5] Evangelos E. Papalexakis, N. D. Sidiropoulos, and R. Bro. From kmeans to higherway coclustering: multilinear decomposition with sparse latent factors. IEEE Transactions on Signal Processing, 2013.
 [6] S. Fernandes, H. FanaeeT, and J. Gama. Dynamic graph summarization: a tensor decomposition approach. Data Mining and Knowledge Discovery, 32(5), 2018.
 [7] A. Gorovits, E. Gujral, E. E. Papalexakis, and P. Bogdanov. Larc: Learning activityregularized overlapping communities across time. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2018.
 [8] E. Gujral and E. E. Papalexakis. Smacd: Semisupervised multiaspect community detection. In Proceedings of the 2018 SIAM International Conference on Data Mining. SIAM, 2018.
 [9] R. Harshman. Foundations of the parafac procedure: Models and conditions for an” explanatory” multimodal factor analysis. 1970.
 [10] T. Kolda and B. Bader. Tensor decompositions and applications. SIAM review, 51(3), 2009.
 [11] Y. Koren. Collaborative filtering with temporal dynamics. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2009.
 [12] J. Leskovec, J. Kleinberg, and C. Faloutsos. Graph evolution: Densification and shrinking diameters. ACM Transactions on Knowledge Discovery from Data (TKDD), 1(1):2, 2007.
 [13] E. Papalexakis, C. Faloutsos, and N. Sidiropoulos. Tensors for data mining and data fusion: Models, applications, and scalable algorithms. ACM Trans. on Intelligent Systems and Technology.
 [14] E. E. Papalexakis. Automatic unsupervised tensor mining with quality assessment. In Proceedings of the 2016 SIAM International Conference on Data Mining, pages 711–719. SIAM, 2016.

[15]
R. Pasricha, E. Gujral, and E. E. Papalexakis.
Identifying and alleviating concept drift in streaming tensor
decomposition.
In
Joint European Conference on Machine Learning and Knowledge Discovery in Databases
, pages 327–343. Springer, 2018.  [16] N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalexakis, and C. Faloutsos. Tensor decomposition for signal processing and machine learning. IEEE Transactions on Signal Processing.
 [17] S. Smith, J. W. Choi, J. Li, R. Vuduc, J. Park, X. Liu, and G. Karypis. FROSTT: The formidable repository of open sparse tensors and tools, 2017.
 [18] S. Soundarajan, A. Tamersoy, E. B. Khalil, T. EliassiRad, D. H. Chau, B. Gallagher, and K. Roundy. Generating graph snapshots from streaming edge data. In Proceedings of the 25th International Conference Companion on World Wide Web, 2016.
 [19] R. Sulo, T. BergerWolf, and R. Grossman. Meaningful selection of temporal resolution for dynamic networks. In Proceedings of the Eighth Workshop on Mining and Learning with Graphs, MLG ’10. ACM, 2010.
 [20] J. Sun, C. Faloutsos, C. Faloutsos, S. Papadimitriou, and P. S. Yu. Graphscope: parameterfree mining of large timeevolving graphs. In Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2007.
 [21] Y. Zheng, L. Capra, O. Wolfson, and H. Yang. Urban computing: concepts, methodologies, and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 5(3):38, 2014.
Comments
There are no comments yet.