As financial services (FS) companies have experienced drastic technology driven changes, the availability of new data streams provides the opportunity for more comprehensive customer understanding. Researchers propose Dynamic Customer Em- beddings (DCE), a framework that leverages customers’ digital activity and a wide range of financial context to learn dense representations of customers in the FS industry. Their method examines customer actions and pageviews within a mobile or web digital session, the sequencing of the sessions themselves, and snapshots of common financial features across our organization at the time of login. They test their customer embeddings using real world data in three prediction problems: 1) the intent of a customer in their next digital session, 2) the probability of a customer calling the call centers after a session, and 3) the probability of a digital session to be fraudulent. DCE showed performance lift in all three downstream problems.
. . (1 )i2s1] where is found using gradient descent on the training set. We also include comparisons of four relevant recurrent recommendation architectures in Table 4. We nd that all RNN architectures outperform the baselines and DCE shows an improvement over recent methods. We (a) Session embedding architecture (b) Customer embedding architecture Figure 1. The DCE framework combines digital session embeddings of click-stream actions, multiple time representations, and an aggregation of customer context up until the current session. hypothesize that changes in the nancial state of customers is an important element inuencing future sessions. Our separate sequential modeling treatment of nancial context may be able to capture these dynamics more effectively than previous methods that either concatenate this information in a single LSTM or fuse with the hidden state.
id: 64a144e9cc2d61d838cd897cab9de131 - page: 3
Table 5 shows the 16 primary customer intents. Each session can have multiple intents, so we treat this as a multi-label classication problem and evaluate predictions with a macro AUROC score across intent classes. Any session without a clear intent defaults to Account Summary. Table 4. Average cosine distance between predicted and actual session embedding Model Cosine Distance Previous session embedding Average session embedding Exponential Moving Average Time LSTM (Zhu et al., 2017) LatentCross (Beutel et al., 2018) JODIE (Kumar et al., 2019) CoSeRNN (Hansen et al., 2020) DCE 0.2107 0.1427 0.1248 0.1176 0.1179 0.1185 0.1185 0.1168
id: cdcec3612f3d2e3f8afdde264ed0168c - page: 4
We t a classier on DCE to compare to other internal baselines detailed in Table 6. We include the previous intent model which only utilized the contextual features (Context only). We also train a vanilla LSTM instead of using the ve separate LSTM modules in DCE. DCE outperforms both of these baselines by almost 10 percent. In DCE+C we re-append context with the DCE embeddings resulting in further improvement. The effectiveness of re-concatenated raw context features in DCE+C demonstrates that this information is still not fully captured by DCE for specic downstream tasks. This supports previous ndings that suggest inclusion of context at multiple points in the pipeline improves performance for recurrent recommendation tasks (Smirnova & Vasile, 2017).
id: 13d5bd1ab0dfcda77d30924e2107871a - page: 4
6.1. Customer Intent Prediction Predicting customer intents in a digital session is an important step in anticipating user needs and providing a seamless customer experience in digital platforms. In this section we evaluate our method on a customer intent prediction task at 6.2. Call A small subset of sessions on our digital platforms lead to customers calling call centers for further assistance. TriagTable 5. Online Intents Intent Credit Report Deposit Overdraft Settings Bank Transactions Account Summary Transaction Management Statements and Documents Activate Redeem with bank Non Purchase Transaction Alter Production Terms Payment Authorized User Replace Card Checks Account Update Table 6. Macro AUROC for predicting customer intents Model Macro AUROC
id: 4eec1e3558162e43fba5888a9c4a54fc - page: 4