Abstract
Web recommendation services bear great importance in e-commerce, as they aid the user in navigating through the items that are most relevant to her needs. In a typical Web site, long history of previous activities or purchases by the user is rarely available. Hence in most cases, recommenders propose items that are similar to the most recent ones viewed in the current user session. The corresponding task is called session based item-to-item recommendation. For frequent items, it is easy to present item-to-item recommendations by "people who viewed this, also viewed" lists. However, most of the items belong to the long tail, where previous actions are sparsely available. Another difficulty is the so-called cold start problem, when the item has recently appeared and had no time yet to accumulate sufficient number of transactions. In order to recommend a next item in a session in sparse or cold start situations, we also have to incorporate item similarity models. In this paper we describe a probabilistic similarity model based on Random Fields to approximate item-to-item transition probabilities. We give a generative model for the item interactions based on arbitrary distance measures over the items including explicit, implicit ratings and external metadata. The model may change in time to fit better recent events and recommend the next item based on the updated Fisher Information. Our new model outperforms both simple similarity baseline methods and recent item-to-item recommenders, under several different performance metrics and publicly available data sets. We reach significant gains in particular for recommending a new item following a rare item.
Introduction
Most modern recommendation engines rely on collaborative filtering logic, often summarized as "people who viewed this also viewed..." While highly effective for popular items, this approach suffers significantly in the long tail—the vast majority of items that have very few transactions. Furthermore, web sessions often involve anonymous users, leaving the system with no historical profile and only the current item as context.
Existing solutions frequently employ "black box" latent factor embeddings. However, these models can be difficult to tune and often struggle with the cold start problem for new items. This research introduces a framework that reliably recommends the "next item" by combining various similarity measures (e.g., Jaccard or Cosine) and external metadata into a single, robust probabilistic model based on Fisher Information.
Methodology
The proposed framework utilizes Markov Random Fields (MRFs) to create a generative model for item interactions.
- Generative Model: The probability of an item appearing in a session is represented as a Gibbs distribution. The "energy function" is defined by the distance between the current item and a set of "anchor" items—typically the most popular items in the dataset.
- Fisher Information: Instead of raw distances, the model calculates the Fisher Score, which is the gradient of the log-likelihood of the model with respect to its parameters. This represents the "direction" the model parameters would need to move to best fit a specific item.
- Fisher Kernel & Distance: By comparing these scores, the authors derive Fisher Distance (FD) and Fisher Conditional Score (FC). This transforms items into a high-dimensional gradient space where similarities are more meaningful, especially for rare items.
- Multimodal Fusion: A key strength of this approach is its ability to blend different data types (e.g., clickstream data and DBPedia metadata) by simply summing their respective Fisher kernels, avoiding complex manual weight tuning.
Experimental Results
The model was evaluated on four diverse datasets: Netflix, MovieLens, Yahoo! Music, and a Books dataset. To specifically test "cold start" and "long tail" performance, the authors analyzed the least frequent 25% of items.
Performance on the Long Tail
The Fisher-based models (FC and FD) consistently outperformed standard baselines like Cosine similarity, Jaccard similarity, and the Euclidean Item Recommender (EIR). As shown in the figure below, the combined "FC + FD Jaccard" method maintains superior recall even as item support (frequency) varies, demonstrating its effectiveness for rare items.
Robustness and Efficiency
A critical finding is the model's robustness regarding the number of "anchor" items used. The performance, measured by Mean Percentile Rank (MPR), tends to saturate quickly. Using as few as 20 to 50 anchor items provides near-optimal results, making the framework computationally efficient for large-scale real-world applications.
Conclusion
By moving from raw item space to a gradient-based Fisher Information space, this framework provides a mathematically grounded solution for session-based recommendations. It excels in handling sparse data and seamlessly integrating metadata. Future work will focus on personalizing the model by incorporating user-specific parameters into the Fisher Information framework.
Summary updated