Several organizations are attempting to increase personalization when interacting with their customers. From tailored coupons to reward programs and even website designs, organizations are seeking new ways to engage directly with individual customers. Organizations have various channels of interactions with their customers and often use different personalization techniques across those touchpoints. Each mode of communication generates different data types and elicits different behavioral responses from the customers. Within companies that have adopted recommender systems to aid with personalization, the systems are frequently developed for each channel separately. However, a customer's response from one channel can be useful in predicting the customer's response in another channel. By building recommender systems independently of one another, companies are neglecting valuable client information.
Utilizing all types of data from all channels of interaction is necessary for creating more powerful and accurate recommendations. Customer information gained across these channels must be shared to shape a collective understanding of the customer. Due to the various modeling choices available, it is essential that an organization selects the appropriate modeling approach
for each channel and data type. Selecting a suboptimal approach can lead to the recommender system not being able to make use of all the available data appropriately, resulting in an incomplete understanding of the customer’s behavior. When a recommender model finds the right approach for each channel based on all the available data, it will produce recommendations that take into account more context. Models such as these are continuously able to integrate new data in order to accommodate changing customer behaviors and dynamic markets.
The most widely-used recommender system technique in the enterprise is collaborative filtering. However, current models that rely only on collaborative filtering are often less effective than models that consider the recent history of customer-product interactions along with customer and product context. Organizations can significantly improve their personalization attempts by using modern hybrid recommender systems. These hybrid systems primarily focus on leveraging two types of techniques simultaneously: collaborative filtering and content-based filtering. Collaborative filtering techniques focus on learning the preferences of customers and the characteristics of products by analyzing the customer-product interaction. On the other hand, content-based filtering makes recommendations by analyzing a customer’s profile and a product’s known features. In a scenario where there is a lack of customer-product interactions, for example the roll-out of a new product, a model must be able to leverage other data. By combining these approaches, the recommender system is able to build a better picture of the customers and products.
Even when utilizing a hybrid system, organizations may fall short by not leveraging recent advancements in deep learning neural networks. These hybrid systems do not consider the images or text associated with a recommendation, and are thus ignoring a large amount of data when a recommendation is made.
For example, when a customer’s preference for a product largely depends on its look, such as for clothing, there will undeniably be a large amount of image data available to analyze, which can be done using convolutional neural networks. Another example involves having written reviews in an online store for products. These reviews can be leveraged by performing a sentiment analysis using recurrent neural networks. Lastly, recurrent neural networks can also be used to learn from a customer’s recent history of interactions. By learning from a history of interactions, models can more accurately predict a customer’s current state and needs, thus improving the recommendations provided. Deep learning hybrid systems provide the ability to utilize a vast amount of unstructured data types, often neglected by organizations. Recommender systems that use this additional data will certainly have an advantage.
Designing a hybrid system adds substantial complexity as there are a multitude of ways to combine models. Sometimes model output scores will be combined by aggregating these scores and weighting them. Other times one model’s output is fed into another as an additional feature. For example, a product brand affinity model can produce a customer embedding or representation of a customer’s preference for certain products, which can then feed into a predictive model. Deep learning is especially suitable for constructing these embeddings with its ability to discover feature representations from scratch. This allows data scientists to focus more on the design of the hybrid recommender system, rather than spending time on feature engineering.
Although there are important considerations
to be made when modeling channels and designing a hybrid system, these attempts to modernize a recommender system may be broken up into stages. Provided that the recommendation problem is challenging, incremental changes in the recommender system maturity will lead to success at every stage. Starting with a collaborative filtering-based system, corporations can move towards modernizing their personalization campaigns by introducing increasingly complex recommender systems. As changes are made and predictive models are created using modern approaches, the models can then be evaluated on hold out test datasets prior to being deployed for A/B testing. When done correctly, a recommender system that utilizes the techniques mentioned will enable organizations to personalize their customers’ experiences immediately, while continuously improving the recommendations based on all the available information.