Temporal-Item-Frequency-Based User-KNN model, which is a neighbor-based approach designed for Next-Basket Recommendation (NBR)

TIFU-KNN Model Overview

1. Purpose and Approach:
   • Purpose: TIFU-KNN is used for predicting the next set of items a user might purchase, leveraging historical purchase data.
   • Neighbor-based Technique: Unlike model-based methods, which might involve complex statistical or machine learning models, TIFU-KNN uses a simpler neighbor-based approach. This method focuses on finding similar users (neighbors) and making predictions based on their behavior.
2. Addressing Limitations of PIF:
   • Personalized Item Frequency (PIF): While PIF provides useful data on how frequently a user purchases items, it doesn’t differentiate between items purchased with the same frequency. This can make it challenging to determine the best items to recommend.
   • Incorporation of Temporal Dynamics: TIFU-KNN enhances PIF by considering the time aspect of purchases. Items purchased more recently are given more significance than those bought earlier, addressing the issue of equally frequent but temporally dispersed purchases.
3. Decay Rates:
   • Group Decay Rate (rg): This decay rate is applied to groups of items, affecting how the frequency of these items decreases over time.
   • Within (group) Decay Rate (rb): This rate is applied within each group, ensuring that more recent appearances of an item have a higher weight in the final frequency calculation compared to earlier ones.
4. Model Components:
   • User Similarity Calculation:
     • User Vector: The historical purchase records of each user are aggregated into a user vector, which represents the user’s purchasing behavior.
     • Time-Decay Mechanism: To account for the evolving preferences of users, TIFU-KNN assigns more weight to recent purchases using hierarchical time-decayed weights. This means that recent items influence the user vector more than older items.
     • Distance Metric: The Euclidean distance between user vectors is used to calculate similarity. A smaller distance indicates higher similarity, and based on this, the model identifies the k-nearest neighbors for each user.
   • Prediction Phase:
     • Repeated Purchase Component (ut): This is based on the user’s own historical purchases, reflecting the tendency to rebuy items.
     • Collaborative Purchase Component (un): This is derived from the average vector of the k-nearest neighbors, capturing the influence of similar users’ purchase behaviors.
     • Linear Combination for Prediction (P): The final recommendation vector PPP is calculated as a linear combination of the repeated purchase component and the collaborative purchase component. The formula is:
     • 
     • Here, α is a hyper-parameter that controls the balance between the two components. It can be adjusted to give more or less importance to the user’s own purchase history versus the purchase history of similar users.

Conclusion

In the rapidly evolving field of Next-Basket Recommendation (NBR), the challenge lies in accurately predicting the next set of items a user will purchase, leveraging historical purchasing behavior. One innovative approach that has gained traction is the Temporal-Item-Frequency-Based User-KNN (TIFU-KNN) model. This method stands out as a neighbor-based approach, utilizing both temporal and frequency data to enhance prediction accuracy.

TIFU-KNN capitalizes on the core concept of Personalized Item Frequency (PIF), which tracks the frequency of individual items purchased by a user. Unlike traditional methods, which may not account for the temporal aspect of purchases, TIFU-KNN introduces a novel mechanism to address this limitation. It incorporates temporal dynamics by assigning decay rates to item frequencies, ensuring that more recent purchases are given higher significance. This temporal weighting is crucial, as it allows the model to adapt to changing user preferences over time, a common phenomenon in retail and e-commerce.

The model operates on a simple yet effective principle: user similarity calculation. It aggregates the historical records of each user into a user vector, which is adjusted to reflect recent purchases more heavily. The Euclidean distance metric is then employed to identify the k-nearest neighbors, providing a collaborative filtering mechanism that leverages the purchasing behavior of similar users. This process enables the model to recommend not only items that the user frequently buys (repeated purchase pattern) but also items that similar users have purchased (collaborative purchase pattern).

In practical terms, TIFU-KNN’s recommendations are generated through a linear combination of two components: the repeated purchase component and the collaborative purchase component. The balance between these components is controlled by a hyper-parameter, allowing for fine-tuning based on specific application needs. This dual approach ensures that the model can effectively recommend both commonly purchased items and new, potentially interesting products that similar users have discovered.