Modeling Contextual Changesin User Behaviour in Fashion e-Commerce

Ashay Tamhane(B), Sagar Arora, and Deepak Warrier

Myntra, Bengaluru, India{ashay.tamhane,sagar.arora,deepak.warrier}@myntra.com

Abstract. Impulse purchases are quite frequent in fashion e-commerce;browse patterns indicate fluid context changes across diverse producttypes probably due to the lack of a well-defined need at the consumer’send. Data from our fashion e-commerce portal indicate that the finalproduct a person ends-up purchasing is often very different from theinitial product he/she started the session with. We refer to this charac-teristic as a ‘context change’. This feature of fashion e-commerce makesunderstanding and predicting user behaviour quite challenging. Our workattempts to model this characteristic so as to both detect and preemptcontext changes. Our approach employs a deep Gated Recurrent Unit(GRU) over clickstream data. We show that this model captures contextchanges better than other non-sequential baseline models.

1 Introduction

Understanding user behaviour is critical for any e-commerce platform in order topersonalise products and induce the user to convert. This becomes easier if theuser has a well-defined need and exhibits cohesive intent. Unlike other domains,purchases are often impulsive in fashion e-commerce. While some users do visita fashion portal out of their need to purchase specific products, a huge numberof folks are there just to explore and transact once they come across a productof their liking. This results in “fluid” browsing patterns that cut across differentproducts categories. In other words, there are many context changes that happenin a typical user session. This makes modeling the user’s behaviour, and hencepersonalisation complex [16,26].

Consider real user sessions illustrated in Figs. 1 and 2. In the session shown inFig. 1, the user browsed a Shirt, and on the very next click switched to Shorts.Again on the third click, we see a switch back to a Shirt (of a different stylealtogether) and then again a switch back to a few Shorts in the next clicks. Afterconsistently viewing a few Shorts, the user switches to a T-shirt on his final click.Such changing contexts makes modeling the user behaviour quite difficult andsimple heuristics will fall short. For example, if we predicted that the user wouldbrowse Shorts in the final click using majority voting, it would have failed. It isimportant to note that before browsing Shorts, the user also looked at a coupleof Shirts and later changed context to Shorts. This is indicative of the fact thatthe user wasn’t really looking for Shorts from the moment he started browsingc© Springer International Publishing AG 2017J. Kim et al. (Eds.): PAKDD 2017, Part II, LNAI 10235, pp. 539–550, 2017.DOI: 10.1007/978-3-319-57529-2 42

540 A. Tamhane et al.

Fig. 1. Men categories session

Fig. 2. Women categories session

but rather changed context during the session. Similarly, in Fig. 2, the user isseen switching between different article types from women categories.

On Myntra, a leading fashion e-commerce platform in India, daily about1.3 million sessions have clicks on at least one product1. Only 35% of thesesessions have a unique product category (considering sessions with more thanone click). We also note that in 41% of the cases the final product a person endsup purchasing is from a different product category than what the user startedthe session with. This further establishes that context changes are common.

(a) ‘Men-Shirts@458’

(b) ‘Women-Kurtas@471’

Fig. 3. Samples from product groups, coded as <category@group number>

Our work attempts to model such impulsive user behaviour so as to bothdetect and preempt change in context. Our approach employs a deep GatedRecurrent Unit (GRU) over the user sessions data (clickstream). Our datasetpresents another challenging scenario of catalogue being highly dynamic due tothe ephemeral nature of fashion products. In such a scenario, training any model1 Here sessions are all the products clicked by a user within a 30 min window.

Modeling Contextual Changes in User Behaviour in Fashion e-Commerce 541

directly at the product level is difficult due to data sparsity. In addition thediversity of a catalogue implies that context changes could happen at a moregranular level even within a product category - people could change contextfrom a graphic T-Shirt to a Polo. For both these reasons we need to work witha higher and stable abstraction of products. To tackle this problem, we leverageour earlier work [21] using word embeddings to create abstractions of productscalled product groups which represent homogeneous product clusters servingsimilar fashion contexts, as illustrated by Fig. 3. Each product in a session isreplaced by its corresponding product group and the context change is modelledacross products groups - both inter and intra category. Hereafter, we will usethe terms product groups, product clusters and contexts interchangeably. Weuse anonymised user sessions data from our fashion portal for our experiments.We evaluate our trained model against two baseline algorithms - majority votingand “Product Group Graph” model. We show in Sect. 4 that GRU outperformsboth these baseline methods.

2 Related Work

There is substantial research on understanding behaviour of online users byanalysing their clickstream data. There are several studies that visualise click-stream data in an attempt to make the complex sequential data more intuitiveto understand [2,14,26–28,30]. Others use cluster based approaches [23] to dif-ferentiate and understand different browsing patterns. There have also beenqualitative studies to observe behavioural differences of buyer and non-buyersessions [22]. While visualising and clustering approaches are directed towardsdescriptive analysis, there is also substantial amount of literature for predictiveanalysis. Studies have attempted to determine the intent of users on e-commercewebsites in both offline [11,12,16,19,20,24] and real time settings [6]. There hasbeen work [8,19] to distinguish user sessions according to the purpose of thevisit as “buyer” or “window shopper”. While much of this literature focusses onpredicting if a user has intent to purchase, there seems to be relatively less focuson generating recommendations for users with multiple context changes withina session.

Modeling user behaviour and generating recommendations is common, espe-cially in studies relevant to search engines. There has been work on recommend-ing web pages in general [1,7,29] and also in specific cases like personalised newsrecommendation [15]. Several modeling approaches have been evaluated in past[10] for such web-personalisation. In this work, we propose modeling the userbrowsing behaviour with a GRU [3,18] for predicting the next user click. Ourapproach is similar to the recent work in [9] in the sense that we propose RNNsfor predicting the next click. However, as mentioned earlier, our dataset presentsa different challenge due to the dynamic catalogue leading to sparsity of styles.In the next section, we describe how we overcome this challenge by using productgroups.

542 A. Tamhane et al.

3 Modeling User Behaviour

As discussed earlier, we wish to model user behaviour that is often prone tocontext changing. Predicting user behaviour is especially useful for use cases likepersonalisation, where we try to identify products that will be of most interest tothe user next. Therefore, the goal is to predict the user’s next click, while makingsure that the recent context of the browsing sequence is taken into account.Specifically, we want to predict the user’s last clicked product group, given restof the product groups he has seen in the session so far. We use three differentmodels for this purpose that we will compare against each other in Sect. 4. Inthis section, we describe these models.

3.1 Majority Voting

In majority voting, we simply predict the next click based on the frequencyof product groups in the session. As an example, consider the same sessionthat we show in Fig. 1 - ‘Men-Shirts@458’, ‘Men-Shorts@20’, ‘Men-Shirts@380’,‘Men-Shorts@20’, ‘Men-Shorts@20’, ‘Men-Shorts@20’, ‘Men-Shorts@20’, ‘Men-Shorts@20’, ‘Men-Shorts@20’, ‘Men-Tshirts@788’. The predictions for nextclick for this session would be in the following order - Men-Shorts@20, Men-Shirts@458, Men-Shirts@380, Men-Tshirts@788. If we need more recommenda-tions, we randomly pick from the top 50 product groups for that gender.

3.2 Product Group Graph (PG Graph)

We now describe our second baseline approach, that uses a product group graphfor prediction. As earlier mentioned, we leverage our earlier work [21] to cre-ate a smaller number of time invariant product groups from the originally largenumber of products (there were 408,155 products in the catalogue for Men and625,171 products for Women at the time of this work). Unlike a typical approachof using product as a document, we consider each session as a document; thewords in the documents correspond to the attributes for all styles observed inthat session. We then train a Word2Vec [17] model on this dataset to learn avector representation of attributes. We aggregate these vector representations toarrive at vector representations of products, sessions (contexts) and users. Thesevectors are used for creating the final clusters, which we refer to as productgroups. These product group embeddings are the centroid of products’ embed-dings. This results in 2813 homogenous and time invariant product groups forMen and 4763 product groups for Women.

From the training set, we build a directional graph where nodes are theproduct groups and edges capture the transition probabilities between productgroups (normalised frequency counts) in all the sessions. We now use a bagof words approach to generate recommendations based on this graph. Specifi-cally, for every click in a session, we get the transition probabilities to all the

Modeling Contextual Changes in User Behaviour in Fashion e-Commerce 543

product groups from the graph. We then aggregate the transition probabili-ties incoming towards every product group and then normalise them. Contin-uing with the same example of Fig. 1, the predictions from this approach arein the following order - ‘Men-Shorts@20’, ‘Men-Shirts@458’, ‘Men-Shirts@380’,‘Men-Shorts@19’, ‘Men-Shirts@469’. Notice that ‘Men-Shorts@19’ and ‘Men-Shirts@469’ are now included in the recommendations even though the user didnot browse them, as they have high transition frequencies from the browsedproduct groups as per the graph.

3.3 GRU

A major drawback of the above two models is that they do not handle sequentialdata. Taking into account the sequential information is critical to our data as itprovides the context to the session. This sequential context becomes even morecritical for sessions with longer duration, as we shall show in Sect. 4. GRUs arevariants of RNNs, that are designed to model sequential data and have beenshown to do this effectively [3,18].

Let xt be the input in the sequence at time t. Let st be the hidden state atstep t. This state acts as the “memory” of the network as it communicates theinformation the network has thus far to the next step. st is calculated based onst−1 and input at time t using Eq. 1. ot is the output at step t, and is calculatedusing Eq. 2. The network can be visualised as shown in Fig. 4 [13]. Note that theparameters U , V and W remain same throughout all the steps.

st = f(U ∗ xt + W ∗ st−1) (1)

ot = softmax(V ∗ st) (2)

Instead of using Eq. 1 that is used by RNNs, GRUs calculate the hidden statest, using Eqs. 3, 4, 5 and 6 that can be visualised by logical gates as shown inFig. 5 [4].

z = σ(xt ∗ Uz + st−1 ∗ W z) (3)

r = σ(xt ∗ Ur + st−1 ∗ W r) (4)

h = tanh(xt ∗ Uh + (st−1 ∗ r) ∗ Wh) (5)

st = (1 − z) ∗ h + z ∗ st−1 (6)

Fig. 4. Unfolding of RNN.

544 A. Tamhane et al.

Fig. 5. GRU gating

We used a 2 layer deep GRU network for all our experiments. While trainingour model, we cross validated on various number of hidden units (50,100,200,500)and found 100 to be the best. We used rmsprop [5] for optimisation. Rmspropused a learning rate of 0.001 and a decay rate of 0.9. We also tried SGD butfound rmsprop to work better. In order to get a visual intuition of this deep learntmodel, we generate sample sessions from the GRU in terms of the product groupsand later randomly sample a product corresponding to each product group in thesession. Figure 6 shows artificially generated samples as “imagined” by the GRU.As can be seen in Fig. 6, the men session starts with our popular value brands‘Roadster’ and ‘Mast & Harbour’. There is a context change to Levis in middle,which is a premium brand and then again the session switches back to the valuebrands. This is typical user behaviour on our platform where users make anaspirational click on a high value item, but later often purchase a lower valueproduct. Another interesting trend which the model seems to capture well canbe observed in the women session shown in Fig. 6. Here, the session changes con-text from western categories (Tshirts, Tops) to Indian ethnic categories (Kurta,Churidar Kurta). The intuition here is that if a user keeps browsing westerncategories, they are likely to eventually move towards Indian ethnic categories.Note that the artificial session for women actually had ten more products at thebeginning, all belonging to western categories that we have truncated for thesake of brevity.

We show quantitative results in Sect. 4. Here, we show specific qualitativeresults for the sessions corresponding to Fig. 1 that were discussed earlier. Ascan be seen in Fig. 7, both GRU and PG Graph place shorts on number oneranking since the user has been clicking a lot of shorts. However, note that GRUseems to have detected the constant change in context of user and hence predictsdifferent article types where as PG Graph suggests more or less similar productsto what the user has been browsing. Note that the last product clicked by theuser was a stripped T-shirt, which GRU predicted at number four in the list.

4 Experiments

We leverage the clickstream data store from our e-commerce platform for allour experiments in this paper. Specifically, for every session browsed on our

Modeling Contextual Changes in User Behaviour in Fashion e-Commerce 545

(a) Men Categories Session

(b) Women Categories Session

Fig. 6. Artificially generated sessions.

e-Commerce portal, we have the sequence of products clicked in the sessionby the user. If we compute the Jaccard Similarity between the sets of browsedproducts’ categories and bought products’ categories across these sessions, about24% sessions have Jaccard similarity less than equal to 0.3 as shown in Fig. 8.The lower values of Jaccard similarity mean that difference between browsedcategories and eventually bought categories is significant. This further establishesthat context changing in our data is common, as described in Sect. 1.

We train separate models for Men and Women and hence firstly we filter outsessions that have both genders present. We then split the rest of the sessionsinto ‘Men’ and ‘Women’ sessions. This results in 7,010,289 Men sessions and9,543,282 Women sessions that are spread across 45 categories for Men and 50

(a) Predictions From GRU

(b) Predictions From PG Graph

Fig. 7. Predictions corresponding to session in Fig. 1

546 A. Tamhane et al.

Fig. 8. Men categories session

categories for Women. As mentioned earlier, the product catalogue is dynamicand the products (or styles) available at any given point in time varies a lot.To tackle this problem, we replace each product by it’s corresponding productgroup.

We use the following notations hereafter. Let Q be the set of all ses-sions/queries with a sequence of product groups <P1, P2, .....Pn> (Pn beingthe true value to be predicted). All models described above output a sequence(sorted in descending order by the probabilities) R of length k <R1, R2...Rk>; kbeing the number of recommendations per query/session. To evaluate the modelsdescribed in Sect. 3, we use the following metrics:

Mean Reciprocal Rank (MRR): The reciprocal rank of a query responseis the multiplicative inverse of the rank of the first correct answer. The meanreciprocal rank is the average of the reciprocal ranks [25] of results for a sampleof queries Q as shown in Eq. 7, where ranki refers to rank of Pn in R.

MRR =1

|Q||Q|∑

n=1

1ranki

(7)

It is worth mentioning that in this particular evaluation metric, we just con-sider one true value (Pn) and hence MRR and Mean Average Precision (MAP)are equivalent.

Normalised Discounted Cumulative Gain (NDCG): For the formulationof NDCG, we assume the true values to be a sequence of k values: Pn and k − 1product groups most similar to Pn. Also, the relevance score in the formula isthe cosine similarity between Pn and other product group Pj . We use the truesequence for formulation of IDCG (rel1 = 1 since sim(Pn, Pn) = 1). For the DCGformulation, we assume the similarity to be 0 if predicted product group does notbelong to the article type of Pn (because embeddings were learnt separately foreach article type); else relevance equals cosine similarity. It is worth mentioningthat this formulation is likely to give quite low absolute values of NDCG sinceour sessions are quite noisy consisting of multiple article types; but the trueideal ranking just considers the similar product groups from the same article

Modeling Contextual Changes in User Behaviour in Fashion e-Commerce 547

type (Pn ’s article type) and hence high IDCG values. Equations 8 and 9 showthe standard NDCG metric.

DCGp = reli +p∑

i=2

relilog2i

(8)

NDCGp =DCGp

IDCGp(9)

Recall (R): We define recall@K as fraction of sessions wherein the true valuePn is retrieved in K recommendations.

Tables 1 and 2 show MRR and NDCG for different values of K (number ofrecommendations) for men and women respectively. The tuples in bracket showGRU improvement over PG-PG graph and Majority vote respectively. As canbe clearly seen, GRU performs significantly better than Majority voting andPG-PG graph, specifically when K increases.

We also experimented by predicting just the category of last click (by ignoringthe predicted product group). This showed a similar improvement in MRR. Forinstance, we report MRR for different values of K for Men and Women in Table 3below. The subscript tuples show GRU improvement over PG-PG graph andMajority vote respectively.

We are especially interested in understanding how GRU performs for sessionswith a high number of context changes. To evaluate this, we consider 20,853 menand 22,576 women sessions with more than two categories, and ones that differin categories of the first and last click in the session. This was done to capturesessions where user started with a particular category, but ended on a totallydifferent category (not just product group). As seen in Table 4, GRU gives abetter recall than the other two baselines for these sessions.

Table 1. MRR and NDCG as a function of K for Men

GRU PG graph Majority voting

MRR NDCG MRR NDCG MRR NDCG

K = 3 0.32(+15.16%,+10.5%) 0.20(+26%,+1.9%) 0.28 0.16 0.29 0.20

K = 5 0.34(+12.17%,+8.7%) 0.18(+17.3%,+1.2%) 0.30 0.16 0.31 0.18

K = 10 0.35(+10.34%,+10.51%) 0.19(+13.87%,+44.49%) 0.31 0.17 0.31 0.13

Table 2. MRR and NDCG as a function of K for Women

GRU PG graph Majority voting

MRR NDCG MRR NDCG MRR NDCG

K = 3 0.31(+17.94%,+9.5%) 0.28(+0.99%,+304%) 0.26 0.28 0.28 0.07

K = 5 0.32(+15%,+7.6%) 0.28(+0.3%,+123%) 0.28 0.28 0.30 0.12

K = 10 0.34(+12.72%,+9.5%) 0.26(+13.64%,+144%) 0.30 0.23 0.31 0.11

548 A. Tamhane et al.

Table 3. MRR for predicting on the category level for different values of K

GRU PG graph Majority voting

Men Women Men Women Men Women

K = 3 0.65(+20.44%,+9.9%) 0.62(+13.06%,+3.8%) 0.54 0.55 0.59 0.6

K = 5 0.66(+19.11%,+8.7%) 0.63(+12.45%,+3.17%) 0.55 0.56 0.6 0.61

K = 10 0.67(+19.17%,+9.6%) 0.64(+12.07%,+3.5%) 0.56 0.57 0.61 0.62

Table 4. Recall for different values of K

GRU PG graph Majority voting

Men Women Men Women Men Women

K = 3 0.32(+20.03%,+12.24%) 0.29(+29.79%,+11.5%) 0.26 0.22 0.28 0.26

K = 5 0.37(+9.22%,+3.01%) 0.34(+16.8%,+0.8%) 0.34 0.29 0.36 0.34

K = 10 0.44(+0.6%,+11.07%) 0.42(+5.12%,+8.33%) 0.44 0.4 0.4 0.39

(a) Men (b) Women

Fig. 9. MRR v/s session length

In Figs. 9a and b, we show how MRR varies as the session length increases(keeping K constant at 20). Clearly, GRUs handle the context change quite wellin the sessions, and perform significantly better for longer sessions.

5 Conclusion

In this work, we attempted to model the behaviour of users in fashion e-commerce, that consists of multiple context changes while browsing. We builtthree models - majority vote, product group graph and GRU for predicting thenext click of users. Through multiple metrics, we showed that GRUs outperformthe other models. We also showed that as the session grows bigger in length,GRUs perform better. This suggests that with more available context, the per-formance of GRU improves.

Modeling Contextual Changes in User Behaviour in Fashion e-Commerce 549

References

1. Agichtein, E., Brill, E., Dumais, S., Ragno, R.: Learning user interaction models forpredicting web search result preferences. In: Proceedings of the 29th Annual Inter-national ACM SIGIR Conference on Research and Development in InformationRetrieval, pp. 3–10. ACM (2006)

2. Cadez, I., Heckerman, D., Meek, C., Smyth, P., White, S.: Model-based clusteringand visualization of navigation patterns on a web site. Data Min. Knowl. Disc.7(4), 399–424 (2003)

3. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recur-rent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)

4. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recur-rent neural networks on sequence modeling. CoRR abs/1412.3555 (2014). http://arxiv.org/abs/1412.3555

5. Dauphin, Y.N., de Vries, H., Chung, J., Bengio, Y.: Rmsprop and equilibratedadaptive learning rates for non-convex optimization. CoRR abs/1502.04390 (2015).http://arxiv.org/abs/1502.04390

6. Ding, A.W., Li, S., Chatterjee, P.: Learning user real-time intent for opti-mal dynamic web page transformation. Inf. Syst. Res. 26(2), 339–359 (2015).http://dx.doi.org/10.1287/isre.2015.0568

7. Gunduz, S., Ozsu, M.T.: A web page prediction model based on click-stream treerepresentation of user behavior. In: Proceedings of the Ninth ACM SIGKDD Inter-national Conference on Knowledge Discovery and Data Mining, pp. 535–540. ACM(2003)

8. Guo, Q., Agichtein, E.: Ready to buy or just browsing?: detecting web searchergoals from interaction data. In: Proceedings of the 33rd International ACM SIGIRConference on Research and Development in Information Retrieval, pp. 130–137.ACM (2010)

9. Hidasi, B., Karatzoglou, A., Baltrunas, L., Tikk, D.: Session-based recommenda-tions with recurrent neural networks. CoRR abs/1511.06939 (2015). http://arxiv.org/abs/1511.06939

10. Kim, D.H., Atluri, V., Bieber, M., Adam, N., Yesha, Y.: A clickstream-based col-laborative filtering personalization model: towards a better performance. In: Pro-ceedings of the WIDM 2004 6th Annual ACM International Workshop on WebInformation and Data Management, pp. 88–95. ACM, NY, USA (2004). http://doi.acm.org/10.1145/1031453.1031470

11. Kim, E., Kim, W., Lee, Y.: Combination of multiple classifiers for the customer’spurchase behavior prediction. Decis. Support Syst. 34(2), 167–175 (2003)

12. Lakshminarayan, C., Kosuru, R., Hsu, M.: Modeling complex clickstream data bystochastic models: theory and methods. In: Proceedings of the 25th InternationalConference Companion on World Wide Web. WWW 2016 Companion, Interna-tional World Wide Web Conferences Steering Committee, Republic and Cantonof Geneva, Switzerland, pp. 879–884 (2016). http://dx.doi.org/10.1145/2872518.2891070

13. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521(7553), 436–444(2015). http://dx.doi.org/10.1038/nature14539

14. Lee, J., Podlaseck, M., Schonberg, E., Hoch, R.: Visualization and analysis ofclickstream data of online stores for understanding web merchandising. In: Kohavi,R., Provost, F. (eds.) Applications of Data Mining to Electronic Commerce, pp.59–84. Springer, New York (2001)

550 A. Tamhane et al.

15. Li, Z., Shang, W.: Personalized news recommendation based on links of web. In:2015 IEEE/ACIS 14th International Conference on Computer and InformationScience (ICIS), pp. 581–584. IEEE (2015)

16. Lo, C., Frankowski, D., Leskovec, J.: Understanding behaviors that lead to pur-chasing: a case study of pinterest. In: Proceedings of the 22nd ACM SIGKDDInternational Conference on Knowledge Discovery and Data Mining, KDD 2016,NY, USA, pp. 531–540 (2016). http://doi.acm.org/10.1145/2939672.2939729

17. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word rep-resentations in vector space. CoRR abs/1301.3781 (2013). http://arxiv.org/abs/1301.3781

18. Mikolov, T., Karafiat, M., Burget, L., Cernocky, J., Khudanpur, S.: Recurrentneural network based language model. In: 11th Annual Conference of the Inter-national Speech Communication Association. INTERSPEECH 2010, Makuhari,Chiba, Japan, 26–30 September 2010, pp. 1045–1048 (2010)

19. Moe, W.W.: Buying, searching, or browsing: differentiating between online shop-pers using in-store navigational clickstream. J. Consum. Psychol. 13(1), 29–39(2003)

20. Montgomery, A.L., Li, S., Srinivasan, K., Liechty, J.C.: Modeling online browsingand path analysis using clickstream data. Mark. Sci. 23, 579–595 (2004)

21. Sagar Arora, D.W.: Decoding fashion contexts using word embeddings. MachineLearning Meets Fashion, KDD 2016 Workshop (2016)

22. Senecal, S., Kalczynski, P.J., Nantel, J.: Consumers’ decision-making process andtheir online shopping behavior: a clickstream analysis. J. Bus. Res. 58(11), 1599–1608 (2005)

23. Su, Q., Chen, L.: A method for discovering clusters of e-commerce interest patternsusing click-stream data. Electron. Commer. Res. Appl. 14(1), 1–13 (2015)

24. Vieira, A.: Predicting online user behaviour using deep learning algorithms. arXivpreprint arXiv:1511.06247 (2015)

25. Voorhees, E.M.: The TREC question answering track. Nat. Lang. Eng. 7(4), 361–378 (2001). http://dx.doi.org/10.1017/S1351324901002789

26. Wang, G., Zhang, X., Tang, S., Zheng, H., Zhao, B.Y.: Unsupervised clickstreamclustering for user behavior analysis. In: Proceedings of the 2016 CHI Conference onHuman Factors in Computing Systems, CHI 2016, NY, USA, pp. 225–236 (2016).http://doi.acm.org/10.1145/2858036.2858107

27. Waterson, S.J., Hong, J.I., Sohn, T., Landay, J.A., Heer, J., Matthews, T.: Whatdid they do? understanding clickstreams with the webquilt visualization system.In: Proceedings of the Working Conference on Advanced Visual Interfaces, pp.94–102. ACM (2002)

28. Wei, J., Shen, Z., Sundaresan, N., Ma, K.L.: Visual cluster exploration of webclickstream data. In: 2012 IEEE Conference on Visual Analytics Science and Tech-nology (VAST), pp. 3–12. IEEE (2012)

29. Zhang, M., Chen, G., Wei, Q.: Discovering consumers’ purchase intentions basedon mobile search behaviors. Flexible Query Answering Systems 2015. AISC, vol.400, pp. 15–28. Springer, Cham (2016). doi:10.1007/978-3-319-26154-6 2

30. Zhao, J., Liu, Z., Dontcheva, M., Hertzmann, A., Wilson, A.: MatrixWave: Visualcomparison of event sequence data. In: Proceedings of the 33rd Annual ACM Con-ference on Human Factors in Computing Systems, pp. 259–268. ACM (2015)