We compare two machine learning approaches for early prediction of shoppers' behaviors, leveraging features from clickstream data generated during live shopping sessions. Our baseline is a mixture of Markov models to predict three outcomes: purchase, abandoned shopping cart, and browsing-only. We then experiment with a mixture of Recurrent Neural Networks. When sequences are truncated to 75% of their length, a relatively small feature set predicts purchase with an F-measure of 0.80 and browsing-only with an F-measure of 0.98. We also investigate an entropy-based decision procedure.
• Computing methodologies → Neural networks; • Applied computing → Online shopping;
Recent e-commerce forecast analysis estimates that more than 1.77
billion users will shop online by the end of 2017 [
Consumers shopping on e-commerce web sites are in uenced
by numerous factors and may decide to stop the current session,
leaving products in their shopping carts. Once a user has interacted
with a shopping cart, abandonment rates range between 25% and
88%, signi cantly reducing merchants’ selling opportunities [
Our baseline approach is based on the latter work, whereas our
new approach uses a mixture of RNNs. Although Wu et al. [
Finally, our approach di ers from most others in its use of a ternary classi cation scheme. We classify clickstreams as either being a purchase, abandon or browsing-only session instead of just purchase and non-purchase.
We consider clickstream data collected over two weeks and consisting of n0 = 1; 560; 830 sessions. A session Si ; i = 1; 2; : : : ; n0, is a chronological sequence of recorded page views, or “clicks.” Let Vi; j be the jth page view of session i so that Si = (Vi;1; Vi;2; : : : Vi;mi ), and mi is de ned as the length of session i. We exclude session i from our experiments if mi < 4, after which n = 198; 936 sessions remain. Sessions with purchases are truncated before the rst purchase con rmation. Table 1 summarizes the n sessions.
Our experiments use both the page type and dwell time of Vi; j . The page type of Vi; j , denoted as Pi; j , belongs to one of eight categories, including search pages, product view pages, login pages, etc. The dwell time, Di; j , of Vi; j is the amount of time the user spends viewing the page, and is not available until the (j + 1)th page view. After the jth page view, the clickstream data gathered for session i is given by Si jj = ((Pi;1; Di;0); (Pi;2; Di;1); : : : ; (Pi; j ; Di; j 1)) where Di;0 is unde ned, i.e., Di;0 = ;. To reduce sparsity, dwell times were placed in 8 bins, evenly spaced by percentiles. 4
Our goal is to classify customer behavior into nal decision categories. In particular, clickstream sequences receive one of the following labels: PURCHASE, if the sequence leads to an item purchase; ABANDON, if an item was left in the shopping cart, but there was no purchase; and BROWSING-ONLY, when the shopping cart was not used. The nal two categories can be combined to investigate PURCHASE vs. NON_PURCHASE behavior. In preliminary studies, the ABANDON sequences were much more similar to the PURCHASE sequences than to the BROWSING-ONLY sequences, so having three categories helped account for some of the confusability of the data. Our eventual goal of applying our models in a live system adds a constraint. The classi er must work for incomplete sequences, without using data from the “future”. 4.1
Bertsimas et al. [
P (Si jCi = ω )P (Ci = ω ) P (Ci = ω jSi ) =
Pω 2Ω P (Si jCi = ω )P (Ci = ω ) with the prior, P (Ci = ω ), estimated from counts.
This model ts the problem’s constraints, because likelihoods can be produced from subsequences without using “future” clicks. Although each chain is trained only on click data, separating data by class implicitly conditions them on class.
Taking inspiration from the Automatic Speech Recognition (ASR)
community and similarities to “Language Modeling”, we adapted
some of their more recent techniques to our problem. In preliminary
experiments, 5-grams performed better than shorter chains, so we
used them. Longer chains cause greater sparsity, so we addressed
this with Kneser-Ney smoothing, which performed best in a study
of language modeling techniques [
Taking further inspiration from the ASR community, we replaced
the Markov-chains in our mixtures with RNNs. Earlier language
modeling work used feed-forward arti cial neural networks [
Earlier RNN-based language models used the “simple recurrent
neural network” architecture [
“Simple” RNN-based language models for ASR were
outperformed by RNNs using the Long Short-Term Memory (LSTM)
conguration and “drop-out” [
In LSTM RNNs, some nodes function as “memory cells”. Some
connections retrieve information from them, and others cause them
to forget. The LSTM equations are [
LST M :htl 1;htl 1;ctl 1!htl ;ctl i sigm *...... gof //+////=*...... ssiiggmm //+///T2n;4n .* hhllt 1 +/ , - , sigm /- , t 1 ctl =f ctl 1+i g htl =o t anh (ctl ) where h and c represent hidden states and memory cells, subscripts refer to time, superscripts refer to layers, Tn;m is an a ne transform from Rn to Rm , refers to element-wise multiplication, and sigm and tanh are applied to each element.
Since LSTM RNNs with drop-out worked much better than
Markov chains for ASR, we replaced the Markov chains with them
in our clickstream mixture models. Additional work was
necessary, since our scenario did not exactly match language modeling.
During training, input tokens were still used to predict following
tokens. During testing, however, our goal was the sequence
probabilities. These were calculated from token probabilities present
in intermediate softmax layers in each LSTM model. Due to the
network architecture, “future” events were not used for this. We
used TensorFlow to implement our LSTM RNNs [
We experimented on the dataset described in Section 3 and summarized in Table 1. It was partitioned into an 80% training/20% testing split, using strati ed sampling to retain class proportions.
All RNNs were trained for 10 epochs, using batches of 20 sequences. We tested 16 combinations of the remaining parameters. The number of recurrent layers was 1 or 2, the keep probability was 1 or 0.5, and the hidden state size was 10, 20, 40, or 80. For a particular mixture model, all the RNNs used the same parameter values.
For each model, we evaluated the prediction performance by truncating the page view sequences at di erent lengths. Table 2 shows the results for the mixture of Markov models, and for one of the mixture of RNN trials. Although we tested 16 di erent RNN parameter combinations, results were so similar that we are only reporting on one of them.
Table 2 reports precision, recall, and F1-measure for each speci c sequence outcome when considering 25%, 50%, 75%, and 100% of the total length of the sequence. For instance, when splitting at 50%, the Markov chain model can predict a PURCHASE with a 0:42 precision and 0:11 recall, resulting in an overall F1-measure of 0:17. For the same conditions, the RNN-based model reaches a precision of 0:82 with 0:71 recall and an F1-measure of 0:76. We also report the accuracy when randomly selecting a class based on the prior distribution of the clickstream corpus. RNN mixture components substantially outperform Markov chain components. This is particularly evident from Figure 2, which shows the F1measure by sequence length for the mixtures of MCMs (dotted line) and of RNNs (solid line). Both models monotonically increase performance as the model observes more data from 25% splits to 100%, but the mixture of RNNs has an immediate edge even with the short 25% sequences. The MCMs present similar F1-measures for the majority class (i.e., BROWSING-ONLY), but it is penalized F1-measure by sequence length by the lack of data for the less represented sequences (i.e., 14.7% for ABANDON and 20.7% for PURCHASE). RNNs instead generalize better due to the memory component that can model long-distance dependencies.
Similarly to Bertsimas et al. [
The legend in Figure 3 corresponds to the true nal state labels of the test examples: dashed red lines for ABANDON sequences, dashed and dotted green lines for BROWSING-ONLY sequences, and solid blue lines for PURCHASE sequences. Ideally, the model would score the PURCHASE sequences (solid blue lines) high, and the other sequences low, and the earlier the distinction could be made, the better. Looking at this gure, there does appear to be some level of discrimination between the categories. In general, the BROWSING-ONLY sequences seem more separable from the PURCHASE sequences than the ABANDON sequences, as expected.
Although Table 2 can be used to compare other work [
We presented two models for the real-time, early prediction of shopping session outcomes on an e-commerce platform. We demonstrated that LSTM RNNs generalize better and with less data than high-order Markov chain models used in previous work. Our approach, in addition to distinguishing browsing-only and cart-interaction sessions, can also accurately discriminate between cart abandonment and purchase sessions. Future work will focus on features, single RNN architectures, and decision strategies.
P
P