Cold Start and Content-Based Hybrids

A marketplace lists 40,000 new products this week. Each has a title, a category, a price and photos, and none has a single click. To a factorisation model every one of them is a row of random numbers, because an item's embedding is learned only from gradients that interactions supply. The newest items, often the ones the business most wants to surface, are the ones the model cannot see.

Matrix factorisation identifies this as structural. This concept is about what to do about it.

Three kinds of cold

Item cold start is a new item with no interactions. It is the most tractable, because items usually arrive with rich metadata.

User cold start is a new user with no history. It is harder, because a new user often brings almost no features: a device, a country, a referral source.

System cold start is a new product with neither, where the only option is a content-based or rule-based system until data accumulates.

The distinction matters because the remedies differ. Item cold start is mostly a representation problem. User cold start is mostly an elicitation problem, solved by onboarding questions or by quickly learning from the first few sessions.

Hybrid factorisation: embed the features, not the ID

The cleanest hybrid replaces each item's free vector with a sum of feature vectors. If item \(i\) has feature set \(F_i\) (its ID indicator, category, brand, tags), its representation is

and the same construction gives users a vector \(p_u\) from their features. The score is \(p_u^\top q_i\) plus biases built the same way. When every item's only feature is its own ID, this reduces to ordinary matrix factorisation. When a new item arrives with the tags tent and ultralight, it receives a meaningful vector immediately, because those tag embeddings were trained on every earlier item that carried them. Kula's LightFM formulation reported that this outperforms purely collaborative and purely content-based models in cold-start and sparse settings, and roughly matches plain factorisation when interaction data is plentiful (Kula, 2015, Metadata Embeddings for User and Item Cold-start Recommendations, arXiv:1507.08439).

The idea is old. Schein and colleagues combined content and collaborative signals in one probabilistic model to recommend items nobody had rated, and introduced evaluation metrics specifically for that setting (Schein et al., 2002, Methods and Metrics for Cold-Start Recommendations, SIGIR, doi:10.1145/564376.564421).

Training the model to expect missing history

A deep hybrid takes both a preference input (a learned collaborative vector) and content features. Trained naively, it leans entirely on the preference input, because that is the stronger signal for every warm item it sees, and it then fails when that input is empty.

DropoutNet addresses this in the optimiser rather than the architecture: during training, the preference input is randomly dropped to zero, so the network is forced to learn to reconstruct useful scores from content alone (Volkovs, Yu & Poutanen, 2017, DropoutNet: Addressing Cold Start in Recommender Systems, NeurIPS). The dropout rate becomes a dial between warm-item accuracy and cold-item robustness.

A third family learns a mapping from content features into an already-trained collaborative space, regressing each warm item's factor vector from its metadata. It leaves the warm model untouched, which is attractive operationally, and it can only be as good as the content's ability to predict collaborative position.

Where practitioners disagree

One camp builds a single hybrid model for all items. The other keeps a strong pure-collaborative model for warm items and routes cold items through a separate content-based retriever, graduating them once they cross an interaction threshold. The single model is simpler and avoids a discontinuity at the threshold. The split design protects warm-item quality, which is where most traffic goes, from any compromise made for cold items. Both are defensible, and the right answer depends on how much traffic cold items receive.

When it breaks

Content similarity is not preference. Two phone cases with identical metadata can differ enormously in quality, and only interactions reveal it. Content gets an item into the candidate set; it cannot tell you whether users will like it.

Exploration is still required, and it costs impressions. Suppose a new item is shown 1,000 times at a true click-through rate of 2%. The standard error of the observed rate is \(\sqrt{0.02 \times 0.98 / 1000} \approx 0.0044\), so a 95% interval is roughly 2% plus or minus 0.9 points. A thousand impressions barely distinguishes a 1.5% item from a 2.5% one. Every cold item consumes traffic that a proven item could have used, and that trade is an explicit policy decision.

Feature leakage inflates cold-start evaluation. Features computed after launch, such as early sales rank or review count, are not available at cold start. Evaluating with them measures a warm model under a different name.

Evaluation must hold out whole items. A random interaction split leaves every test item with training interactions, so it measures warm performance. A cold-start claim needs items absent from training entirely.

Popularity bias reasserts itself. Once cold items graduate into the collaborative model, they compete with incumbents that have years of gradient, and without a deliberate boost many never accumulate enough interactions to be recommended at all.