Multi-Business Prediction (MBP) Overview

• MBP is a business-aware prediction structure that integrates a shared user–item encoder with tailored, business-conditioned generative decoders.
• It employs a Mixture-of-Experts approach to fuse context and business information, mitigating interference and enabling cross-business transfer.
• MBP enhances recommendation outcomes by addressing seesaw effects and representation confusion, as validated by empirical ablation studies.

Searching arXiv for the cited papers to ground the article in current metadata. Multi-Business Prediction (MBP) most directly denotes, in recent recommendation research, a business-aware prediction structure that converts a shared user–item representation into business-specific generative predictions so that a single model can serve multiple heterogeneous businesses without forcing them to share an undifferentiated prediction head (Li et al., 3 Apr 2026). The term is not fully standardized across adjacent literatures. In predictive process analytics, a closely aligned pattern appears as generic representation learning followed by task-specific fine-tuning, although the paper does not explicitly coin MBP as a term (Pfeiffer et al., 2021). In protein–ligand binding affinity prediction, the same initials denote “Multi-task Bioassay Pre-training,” a different multi-task pre-training framework built around shared encoders and task-specific heads (Yan et al., 2023).

1. Terminological scope and definitional boundaries

In current arXiv usage, MBP has one explicit meaning and two closely related or acronym-sharing meanings. The explicit meaning is the one introduced in MBGR at Meituan, where MBP is a component inside a larger generative recommendation system rather than the entire system (Li et al., 3 Apr 2026). A common misconception is therefore to treat MBP and MBGR as interchangeable. In the cited formulation, MBGR comprises three key components—Business-aware semantic ID (BID), Multi-Business Prediction (MBP), and Label Dynamic Routing (LDR)—and MBP is specifically the prediction structure layered between a shared sequence representation and business-conditioned token generation (Li et al., 3 Apr 2026).

arXiv id	Expansion or aligned formulation	Domain
(Li et al., 3 Apr 2026)	Multi-Business Prediction	Generative recommendation at Meituan
(Pfeiffer et al., 2021)	Multi-Perspective Process Network, described as conceptually aligned with MBP	Predictive process analytics
(Yan et al., 2023)	Multi-task Bioassay Pre-training	Protein–ligand binding affinity prediction

The process-mining formulation is relevant because it explicitly targets “one representation framework, many business-process prediction tasks and logs,” implemented as self-supervised pre-training of a shared encoder and downstream task-specific heads (Pfeiffer et al., 2021). The bioassay formulation is relevant because it demonstrates a parallel design pattern: heterogeneous supervision, shared-bottom representation learning, and separate heads for label-specific tasks under noisy conditions (Yan et al., 2023). This suggests that MBP, in a broader methodological sense, is best understood as a family of shared-representation, task-conditional prediction schemes rather than a single canonical algorithm.

2. MBP within MBGR

The Meituan formulation arises in a multi-business industrial setting involving several distinct businesses, denoted A–D, including food delivery, various local services, entertainment, and healthcare (Li et al., 3 Apr 2026). Each business has its own item universe, behavior patterns, and business goals, while users interact across businesses, creating both transfer opportunities and interference risks. MBP is introduced precisely at the point where a shared sequential representation must be converted into predictions that remain useful for each business rather than collapsing toward the dominant one (Li et al., 3 Apr 2026).

The overall MBGR pipeline is organized as follows. A user sequence is represented as $$S_u = \{(i_{u,1}, b_{u,1}, t_{u,1}), \ldots, (i_{u,L}, b_{u,L}, t_{u,L})\}$$. Item Semantic IDs (SIDs) are first encoded by BID using token embeddings together with a business embedding to form a business-aware item representation. A Transformer or HSTU backbone then processes the sequence of such representations and produces contextualized representations of user interests over time. MBP sits on top of this shared encoder output: it takes each shared representation $\mathbf{e}$ together with a business embedding $\mathbf{b}$, transforms them into a business-specific representation $\mathbf{e}^b$, and passes that representation into the business-conditioned decoder used by BID in order to generate token predictions for each business (Li et al., 3 Apr 2026).

Formally, MBP uses a parameter-sharing Mixture-of-Experts style transformation. It first fuses context and business information by concatenation, $\mathbf{z}^b = [\mathbf{e}, \mathbf{b}]$. It then computes adaptive expert weights $$\mathbf{g}^b = \operatorname{SiLU}(\operatorname{FFN}_{\text{gate}}(\mathbf{z}^b))$$, and aggregates shared experts as

$$\mathbf{e}^b = \sum_{k=1}^{K} g_k^b \cdot \operatorname{FFN}^{\text{exp}}_k(\mathbf{z}^b)$$

(Li et al., 3 Apr 2026). The important architectural point is that experts are shared across businesses, but gating depends on both representation and business, so different businesses or contexts activate different mixtures of shared capacity.

This arrangement gives MBP a dual role. It reuses knowledge across businesses through shared experts, but it avoids forcing all businesses through the same final mapping. In the MBGR paper, MBP is therefore not the sequence model, not the SID tokenizer, and not the routing mechanism for labels; it is the business-aware prediction layer between shared sequence modeling and business-conditioned token decoding (Li et al., 3 Apr 2026).

3. Training objective, label routing, and supervision density

The MBGR training objective combines a multi-business InfoNCE loss for prediction with a reconstruction loss for BID:

$$\mathcal{L} = \mathcal{L}_{\text{infoNCE}} + \lambda \cdot \mathcal{L}_{\text{recon}}.$$

The predictive term is indexed over businesses, batch items, and token positions, and it includes both business weights $w_b$ and a time-decay weight $w_t$ (Li et al., 3 Apr 2026). The time-decay term is defined as $\mathbf{e}$0, which emphasizes more recent targets. The reconstruction term preserves semantic integrity in BID through token-space reconstruction (Li et al., 3 Apr 2026).

The central difficulty in multi-business sequence supervision is that the immediate next event may belong to only one business, whereas MBP is intended to learn business-specific prediction capabilities for all businesses from the same mixed sequence. LDR resolves this by routing, for each time step $\mathbf{e}$1 and each business $\mathbf{e}$2, the label to the next future interaction of that same business:

$\mathbf{e}$3

If no such future interaction exists, the loss for that business at that position is masked (Li et al., 3 Apr 2026). MBP thus produces a prediction for each $\mathbf{e}$4 pair, and LDR supplies the corresponding business-specific target.

This routing changes the supervision regime from sparse to dense at the business level. Instead of supervising only the business that happens to appear next in the raw sequence, the model receives a separate target for each business whenever future evidence exists. In the paper’s formulation, this is crucial because MBP is designed to learn $\mathbf{e}$5 for multiple businesses simultaneously from interleaved user histories (Li et al., 3 Apr 2026).

A related misconception is to interpret MBP as a purely head-level trick. The training equations show that it is more than that: it participates in a joint optimization in which business-conditioned representation transformation, business-aware decoding, per-business routing, and explicit reweighting all interact. The business weights are a further balancing mechanism; the empirical setting reported in the paper assigns higher weights to smaller or high-value businesses, for example B: 1.5, C: 1.3, A: 0.9, D: 1.0 (Li et al., 3 Apr 2026).

4. Seesaw phenomenon, representation confusion, and the role of MBP

The MBGR paper identifies two critical issues in multi-business generative recommendation. The first is the seesaw phenomenon: with a single global generative head and a unified SID space, improvements for one business degrade another, and large businesses dominate updates. The second is representation confusion: a unified SID space and shared prediction head fail to distinguish distinct semantic information across businesses, causing token-level and hidden-state entanglement (Li et al., 3 Apr 2026).

MBP is introduced to address the prediction-stage part of both problems. BID already injects business context into item representation and token reconstruction, but BID alone does not specify how shared sequence representations should be mapped into business-specific next-item predictions. MBP adds that missing step by applying a business-conditioned MoE transformation before decoding. The consequence is that the same shared representation $\mathbf{e}$6 can be routed to different expert mixtures depending on whether the target business is A, C, or another domain, and the decoded output is optimized for that business rather than for a single average objective (Li et al., 3 Apr 2026).

The seesaw mitigation mechanism has three parts. First, the business-aware MoE means that gradients from one business primarily affect the experts it actually uses rather than forcing all businesses through identical parameters. Second, the business weights $\mathbf{e}$7 counteract imbalance by amplifying the contribution of smaller, fast-changing, or high-value businesses. Third, LDR supplies denser supervision for each business, improving robustness for businesses that would otherwise receive fewer next-event labels in a mixed sequence (Li et al., 3 Apr 2026).

The representation-confusion mitigation mechanism is complementary. BID introduces business-aware encoding and decoding, while MBP applies a business-conditioned transformation before decoding. Together they create what the paper describes as a “disentangled yet shared” structure: shared backbone and codebook for cross-business generalization, but business-aware encoders, decoders, and prediction transformations for specialization (Li et al., 3 Apr 2026). This combination is empirically associated with clearer clustering by business in embedding visualizations and with stronger business-specific recommendation metrics.

5. Empirical behavior, ablations, and operational profile

The MBGR evaluation reports offline and online evidence that MBP materially contributes to multi-business recommendation quality. In generative Hit@10, an HSTU baseline reports All: 0.0214, B: 0.0127, C: 0.0198, whereas HSTU-MBGR reports All: 0.0410, B: 0.0554, C: 0.0398; business A also improves from 0.0231 to 0.0252 (Li et al., 3 Apr 2026). The paper interprets this as evidence that the multi-business architecture lifts smaller businesses without sacrificing the dominant one.

The MBP ablation is especially direct. Removing MBP yields Hit@10 values of All = 0.0330, B = 0.0401, C = 0.0312; the full MBGR system with MBP restores All = 0.0410, B = 0.0554, C = 0.0398 (Li et al., 3 Apr 2026). The weight-balancing ablation is similarly specific: uniform $\mathbf{e}$8 gives All = 0.0368, B = 0.0487, C = 0.0342, while the empirical weighting scheme improves them to All = 0.0410, B = 0.0554, C = 0.0398 (Li et al., 3 Apr 2026). These results isolate MBP as a major contributor to cross-business stabilization and improvement.

Downstream ranking metrics show the same pattern. In CTCVR GAUC, the baseline and MBGR scores are reported as All: 0.7748 to 0.8040, Business B: 0.7594 to 0.8258, Business C: 0.8852 to 0.9052, and Business D: 0.7466 to 0.7717 (Li et al., 3 Apr 2026). Online A/B testing reports CTCVR uplift of +3.0% for A, +7.5% for B, +4.5% for C, +5.2% for D, with a weighted average of +3.98% (Li et al., 3 Apr 2026). Although the online evaluation concerns the full MBGR stack rather than MBP in isolation, the ablation evidence indicates that MBP is central to those gains.

Implementation details reinforce the interpretation of MBP as a lightweight specialization layer rather than a replacement for the backbone. The reported generative training uses PyTorch, 16 Transformer layers, 8 attention heads, and embedding dimension 128. The MBP MoE performs best at $\mathbf{e}$9 experts; $\mathbf{b}$0 yields All Hit@10 = 0.0379, $\mathbf{b}$1 yields 0.0410, and larger values such as 16 or 32 show no further gain and a slight drop attributed to parameter fragmentation (Li et al., 3 Apr 2026). The time-decay parameter $\mathbf{b}$2 is reported as working best. In deployment, MBGR is not used as a full end-to-end retrieve–rank generator, but as a generator of user embeddings and item embeddings that serves as an additional retrieval channel and as features for ranking models; MBP’s computation is described as lightweight relative to the backbone (Li et al., 3 Apr 2026).

6. Related formulations in process mining and bioassay pre-training

A closely aligned formulation appears in predictive process analytics, where the Multi-Perspective Process Network (MPPN) is explicitly described as a generic representation learner for business processes, intended to support “multiple process prediction tasks” over the same event logs through “pre-train once, fine-tune many tasks” (Pfeiffer et al., 2021). The model treats each event-log perspective as a time series, converts each perspective into a Gramian Angular Field image, applies a shared CNN1 across perspectives, pools by concatenation, and uses a two-layer fully connected network NN2 to produce a feature vector $\mathbf{b}$3 (Pfeiffer et al., 2021). Self-supervised multi-task next-event prediction over all attributes in $\mathbf{b}$4 trains the encoder so that $\mathbf{b}$5 can later support next activity prediction, next resource prediction, next timestamp prediction, outcome prediction, clustering, anomaly detection, and case-based retrieval (Pfeiffer et al., 2021). The paper explicitly states that it does not coin “Multi-Business Prediction (MBP)” as a term, but it characterizes the architecture as conceptually aligned with MBP in the sense of one generic representation framework supporting many business-process prediction tasks and logs (Pfeiffer et al., 2021).

The protein–ligand literature uses the same initials differently. “Multi-task Bioassay Pre-training” defines MBP as a supervised pre-training framework that uses a large, noisy, heterogeneous dataset, treats different affinity types as separate prediction tasks, and augments regression with pairwise ranking within each bioassay (Yan et al., 2023). Its architecture is a shared-bottom encoder $\mathbf{b}$6 with four task-specific heads: IC50 regression, K regression, IC50 ranking, and K ranking. The total pre-training objective is $\mathbf{b}$7, and the framework is then fine-tuned on higher-quality downstream benchmarks (Yan et al., 2023). In direct terminological terms, this MBP is not Multi-Business Prediction. Methodologically, however, it reinforces the same pattern of heterogeneous supervision, shared representation learning, and task-specific adaptation under label noise.

Taken together, these neighboring formulations indicate that the most stable cross-domain content of MBP-like methods is architectural rather than nominal. Shared encoders, task- or business-conditioned heads, and auxiliary objectives that preserve transfer while limiting interference recur across recommendation, process mining, and scientific prediction (Pfeiffer et al., 2021, Yan et al., 2023). This suggests that the Meituan MBP formulation is a domain-specific instantiation of a broader research program centered on shared latent spaces with structured specialization.

7. Limitations and prospective directions

The MBGR paper does not present a separate limitations section devoted exclusively to MBP, but it identifies several implicit constraints. Scalability to many more businesses may require more sophisticated expert-sharing strategies; cold-start businesses rely heavily on transfer through shared experts; deployment remains incremental rather than replacing the whole retrieval–ranking stack; and explicit modeling of business relations may improve robustness and interpretability beyond the current MoE gating scheme (Li et al., 3 Apr 2026).

The process-mining formulation highlights a different set of issues that are also relevant to MBP in the broad sense. The authors do not systematically tune hyperparameters per log or task, note the black-box nature of the model, and identify cross-log pre-training and explanation of attribute influence as natural future directions (Pfeiffer et al., 2021). These points matter because business-conditioned prediction systems inherit the same tension between generic transfer and domain-specific interpretability.

The bioassay formulation exposes a further limitation class: noisy labels and approximate input structures can be mitigated but not eliminated by multi-task training and intra-group ranking (Yan et al., 2023). A plausible implication is that multi-business recommendation systems using MBP-like structures may also benefit from explicit noise modeling, more structured pair or list sampling, or hierarchical conditioning mechanisms once the number of businesses or objectives becomes large.

In its strictest contemporary sense, then, Multi-Business Prediction is the business-aware MoE prediction structure inside MBGR that maps shared sequence representations to business-specific generative outputs and is trained jointly with BID and LDR to mitigate seesaw and representation confusion (Li et al., 3 Apr 2026). In the broader methodological landscape, it belongs to a larger class of shared-representation, multi-head, pre-train-and-adapt systems that recur across business-process modeling and scientific prediction, even when the acronym itself refers to something else (Pfeiffer et al., 2021, Yan et al., 2023).